VERITAS Software Corporation

VERITAS File System Performance White Paper

VERITAS Software Corporation 1600 Plymouth Street Mountain View, CA 94043 1.800.258.8649 in the USA 415.335.8000 FAX 415.335.8050 E-mail: [email protected] World Wide Web: http://www.veritas.com/ VERITAS, the VERITAS logo, VxFS, VxVM, FirstWatch and VERITAS FirstWatch are registered trademarks of VERITAS Software Corporation. VxServerSuite and VxSmartSync are trademarks of VERITAS Software Corporation. Other product names mentioned herein may be trademarks and/or registered trademarks of their respective companies. 1996 VERITAS Software Corporation. All rights reserved. 11/96

Table of Contents
Table of Contents......................................................................................................................................i Introduction ..............................................................................................................................................3 Benchmark Testing platform.............................................................................................................4 Benchmark Program - VERITAS vxbench .......................................................................................5 Benchmarks used in this Report .......................................................................................................7 VERITAS File System .............................................................................................................................9 Journaling..........................................................................................................................................9 Performance Components................................................................................................................9 Extent Based Allocation.........................................................................................................................11 Buffered File System Benchmark...................................................................................................11 Buffered Aged File System Benchmark..........................................................................................14 Cache Policies .......................................................................................................................................17 File System Alignment....................................................................................................................18 Direct I/O.........................................................................................................................................18 Discovered Direct I/O......................................................................................................................19 Controller Limited Benchmark - 4 way Stripe .................................................................................19 Disk Limited Benchmark - 8 way Stripe ..........................................................................................22 Tuning UFS File Systems ...............................................................................................................26 Tuning UFS ..............................................................................................................................28 Passing the GB/sec Barrier.............................................................................................................28 Quick I/O for Database ..........................................................................................................................31 File Systems and Databases ..........................................................................................................31 UNIX raw I/O ...................................................................................................................................31 raw I/O Limitations....................................................................................................................32 Quick I/O for Database ...................................................................................................................32 Multiple Thread Random I/O Benchmarks .....................................................................................32 Direct I/O vs. Quick I/O for Database ....................................................................................................35 Direct I/O vs. Quick I/O for Database Benchmarks ........................................................................35 Conclusion .............................................................................................................................................41

VERITAS File System

Performance

VERITAS Software

ii

File System Performance

Chapter

VERITAS Software Corporation

VERITAS File System Performance - White Paper

Introduction
The VERITAS storage management product line has been designed in response to the needs of commercial computing environments. These are the systems that are now supporting mission critical applications, and playing a major role in corporate computing services. Typical environments include online transaction processing systems, both inter as well as intra-networked database servers, and high performance file services. VERITAS specializes in systems storage management technology which encompasses a product line that offers high performance, availability, data integrity, and integrated, online administration. VERITAS provides three complementary products: VERITAS FirstWatch, VERITAS Volume Manager, and the VERITAS File System. VERITAS FirstWatch is a system and application failure monitoring and management system that provides high-availability for mission-critical applications. FirstWatch dramatically increases server application availability through the use of duplicate cluster monitoring processes on each node in a FirstWatch system pair. These monitoring processes communicate with each other over dedicated, duplicated heartbeat links. VERITAS Volume Manager is a virtual disk management system providing features such as mirroring, striping, disk spanning, hot relocation and I/O analysis. The VERITAS Visual Administrator exists as a graphical interface to VERITAS Volume Manager offering visual representation and management of the virtual disk subsystem including drag and drop features and simple or complex device creation. The VERITAS File System is a high availability, high performance, commercial grade file system providing features such as transaction based journaling, fast recovery, extentbased allocation, and on-line administrative operations such as backup, resizing and defragmentation of the file system. This report describes the performance mechanics of the 2.3 version of the VERITAS File System. This report will provide a discussion of the key performance components in the VERITAS File System. This report will also present a series of benchmark tests comparing the throughput and CPU utilization of different VERITAS File System component technologies, as well as provide some tests comparing the VERITAS File System (VxFS) with the Solaris UNIX File System (UFS). VERITAS Software developed a series of benchmark tests for the express purpose of testing performance throughput of the installable file system, software component of the
VERITAS File System

Performance

UNIX OS. Since there are a number of components involved in computer file system, it was deemed necessary to develop a methodology, and later a program, to allow the configuration and running of a number of different I/O streams, in an effort to understand the role that file systems play in the overall file I/O model. The testing included assembling a hardware test platform that would give the testers the ability to induce hardware bottlenecks at some specific points in order to see the overall effect on file system throughput. For the sake of this testing the performance hardware bottleneck areas focused on are: CPU - For most tests the desire was to not create a CPU bottleneck, but rather allow enough system CPU cycles, and system RAM, to exist for testing so that the throughput emphasis could be analyzed elsewhere. I/O Controller - The controller bus utilized in the tests was a Fast/Wide SCSI bus with the theoretical throughput of 20mb/sec. In some tests the controller was saturated in order to determine the overall effect on file system I/O performance. Disk - The disks utilized in the tests were all Fast Wide Differential SCSI hard drives with an aggregate throughput of approximately 6.5mb/sec. In other tests the disks were saturated in order to determine their effect on file system I/O performance.

Benchmark Testing platform

The complete breakdown of the hardware testing platform used for all benchmark tests is as follows: Hardware: SUN Microsystems Ultra 4000 equipped with: 4 Ultra SPARC Processors running at 167.5mhz 256 MB system memory 3 Andataco RSE-254S3w Storage Subsystems each RSE-254S3W contained 8 Seagate ST3255OWD drives each RSE-254S3W contained two fast/wide SCSI busses 6 SUN 1062A F/W/D SBus controllers were used Software: Solaris 2.5.1 VERITAS File System version 2.3 VERITAS Volume Manager version 2.3 VERITAS Software vxbench benchmark program Test Specifications: All measurements were made with VERITAS Volume Manager RAID-0 array - volumes These volumes were configured with 64 KB stripe units and the file system was aligned automatically by the combination of the VERITAS File System and Volume Manager, to stripe unit boundaries
4

VERITAS Software

File System Performance

Benchmark Program - VERITAS vxbench

VERITAS engineering developed a benchmark program specifically to allow a user to create a variety of I/O environments. Using this type of tool would allow the tester the ability to perform a wide variety of performance measurements on the installable file system component of the UNIX OS. The program developed is called vxbench and some of the features include the ability to utilize: All VxFS file system options All UFS file system options raw I/O, UFS, or VxFS partitions Multiple block sizes Multiple file sizes Random and sequential I/O All of the test results described in this report were derived using vxbench. As you will notice from the description below vxbench can perform I/O in utilizing multiple I/O streams. One multiple stream mode is to perform I/O to a single file using multiple threads, indicative of a database type application workload. Another multiple stream mode is to perform I/O to multiple files, via multiple threads, indicative of a multiple user server environment. The following is the list of vxbench options: usage: vxbench -w workload [options] Valid options are: -h -P -p -t -m -s -v -k -M print more detailed help message use processes for users, threads for multithreaded I/O (default) use processes for users and for multithreaded I/O use threads for users and for multithreaded I/O lock I/O buffers in memory for multiuser tests only print summary results for multithreaded tests print per-thread results print throughput in KB/sec (default) print throughput in MB/sec filename ...

-w workload selects a type of I/O workload valid workloads are: read sequential read of the test files write sequential write of the test files rand_read random read of the test files rand_write random write of the test files rand_mixed mix of random reads and writes mmap_read use mmap to read the test files mmap_write use mmap to overwrite the test files -i subopts specify sub options describing test valid sub options are: nrep=n repeat the I/O loop in the test n times
5
File System Performance

VERITAS Software

nthreads=n iosize=n fsync remove iocount=n reserveonly maxfilesize randseed truncup rdpct=n

number of threads accessing each file size of each I/O do an fsync on the file after writing it remove each file after the test number of I/Os reserve space for the file but don't do I/O maximum offset in KB for random I/O tests seed value for random number generator set an initial file size for random I/O set read percentage of job mix for mixed tests

-o opentype specify flags for opening the file valid opentypes are: append use appending writes sync set the O_SYNC flag for synchronous file I/O trunc truncate the test files on open -c cacheopts specify VxFS caching advisories valid cache options are: direct use direct I/O to bypass the kernel cache dsync use data synchronous I/O noreuse set the VX_NOREUSE cache advisory random set the VX_RANDOM cache advisory seq set the VX_SEQ cache advisory -e extsize specify a fixed extent size -r reservation specify space reservation -f flags specify flags for reservation and fixed extents valid flags are: align require aligned extents chgsize set the file size to the reservation size contig require contiguous allocation noextend don't allow writes to extend the file noreserve allocate space but don't set file reservation trim trim reservation to file size on last close

Specifying multiple filenames will run tests in parallel to each file, thus simulating multiple simultaneous users. If multiple threads are also specified, then each simulated user will run multiple threads so the total number of I/O threads will be 'users * nthreads'. An example usage of vxbench would be as follows. If you wanted to measure I/O throughput of sequentially writing a 1024 MB file in 8 KB blocks, you would invoke vxbench as follows: ./vxbench -w write - i iosize=8k,iocount=128k /dev/vx/dsk/perfvol1 There is also a built-in help file that can be invoked by: ./vxbench -h.

VERITAS Software

File System Performance

Benchmarks used in this Report

Each series of benchmarks in this report includes configuration and parameter information that was used during the series of tests. In some cases there was an effort to test the file system software components with default settings, in other cases certain changes were made based upon the testing modes used. As mentioned previously, in some instances the hardware configuration was limited based upon the testing mode utilized. Specific information is available with each series of test results. All benchmark tests run in this report were done using the vxbench program. Also except where noted, all benchmark testing involved testing sequential I/O. Finally, anyone wishing to use vxbench for performance testing of their file systems may obtain the program from VERITAS Software for no charge.

CPU Measurements
A note regarding the CPU measurements reported in this paper. The vxbench program measures two types of CPU times: 1. time spent in the operating system (system time) 2. time spent in the application (user time) The way in which the measurements were reported in these tests was that both times were combined to come up with a single measurement of CPU impact. If the application time was reported as 10.5 seconds and the system time was 189.5 seconds, the final measurements would be reported as 200 CPU seconds. CPU utilization is, strictly speaking, this time divided by the elapsed time (which is not reported). The reason for using CPU seconds is to compare the relative CPU seconds per file system option when transferring the same amount of data.

VERITAS Software

File System Performance

VERITAS Software

File System Performance

Chapter

VERITAS File System

In response to the growing file system needs of commercial computing environments VERITAS Software developed their own installable file system initially for the UNIX environment. The VERITAS File System is a modern file system which is semantically similar to UFS, but which has been redesigned to support server-class file storage requirements. It adds a method (called journaling or intent-logging) to increase reliability, and uses more efficient, extent-based allocation policies as well as layout and caching modifications to more closely meet the I/O requirements of commercial applications and databases.

Journaling
The VERITAS File System employs a variation on the general file system logging or journaling technique by employing a circular intent log. All file system structure changes, or metadata changes, are written to this intent log in a synchronous manner. The file system will then periodically flush these changes out to their actual disk blocks. This increases performance by allowing all metadata writes to be written out to the permanent disk blocks, in an ordered manner out of the intent log. Because the journal is written synchronously, it may also be used to accelerate small (less than or equal to 8KB) synchronous write requests, such as those used for database logging. Writes of this class may be written to the journal, a localized sequential block store, before they are moved to their places in the larger file system; this can reduce head movement and decrease the latency of database writes. By using this intent log, the VERITAS File System can recover from system downtime in a fraction of the time. When the VERITAS File System is restarted in the same scenario, the system simply scans the intent log, noting which file system changes had completed and which had not, and proceed accordingly. In some cases, the VERITAS File System can roll forward changes to the metadata structures, because the changes were saved in the intent log. This adds availability and integrity to the overall file system.

Performance Components
The VERITAS File System has been developed with many of the latest industry file system performance improvements in mind. These improvements can be divided into the following feature categories: Extent Based Allocation

VERITAS File System

Performance

Unlike traditional UNIX file systems, which assign space to files one block at a time, the VERITAS File System allocates blocks in contiguous segments called extents. Extent sizes are chosen based on the I/O pattern of the file, or may be explicitly selected to suit the application. Extent-based allocation can accelerate sequential I/O by reducing seek and rotation time requirements for access, and by enabling drivers to pass larger requests to disks. Cache Policies The UNIX operating system supplies a standard asynchronous mode for writing to files, in which data is written through a write-back page cache in system memory, to accelerate read and write access. It also calls for a synchronous mode, which writes through the cache immediately, flushing all structures to disk. Both of these methods require data to be copied between user process buffer space to kernel buffers before being written to the disk, and copied back out when read. However, if the behavior of all processes that use a file is well-known, the reliability requirements of synchronous I/O may be met using techniques which offer much higher performance, often increasing file access to about the speed of raw disk access. The VERITAS File System provides two types of cache policies which enable these types of performance improvements. The first method is called Direct I/O, and using this method the VERITAS File System does not copy data between user and kernel buffers; instead, it performs file I/O directly into and out of user buffers. This optimization, coupled with very large extents, allows file accesses to operate at raw-disk speed. Direct I/O may be enabled via a program interface, via a mount option, or with the 2.3 version of the VERITAS File System, Direct I/O can be invoked automatically based upon the I/O block size. This feature is known as Discovered Direct I/O. The second cache policy available with the 2.3 version of the VERITAS File System is the Quick I/O for Database. While Direct I/O improves many types of large I/O performance, the single writer lock policy of the UNIX OS creates a performance bottleneck for some types of file system writes. Database application writes are particularly affected by this. Included in the VERITAS ServerSuite Database Edition 1.0, the Quick I/O for Database bypasses the single writer lock policy in the UNIX OS by representing files to applications as character devices. This allows database applications suited to utilizing raw partitions, the ability to operate like they are using a raw partition, on a file system. This combines the manageability of file systems with the performance of raw partitions.

VERITAS Software

10

File System Performance

Chapter

Extent Based Allocation

The VERITAS File System uses the much more efficient extent based allocation that improves the way in which large files are handled. Rather than linking indirect blocks of addresses, the VERITAS File System uses extent addresses which list a starting block address and a size. The disk blocks allocated for a file are stored in contiguous extents starting at the starting block address, and extending contiguously the number of blocks denoted by the size number. Because a single pointer addresses more than one block, an extent-based file system requires fewer pointers and less indirection to access data in large files. UFS, with its 12 direct pointers, can only directly address up to 96KB of data (using 8KB blocks) without requiring at least one extra block of pointers and an indirect access. The VERITAS File System, with its 10 pointers to extents of arbitrary size, can address files of any supported size directly and efficiently. What this translates to, is that when a large file is accessed in the VERITAS File System, the blocks needed can usually be found with no indirection, or directly. This direct addressing ability of the VERITAS File System dramatically increases the performance when the file system handles large files. The VERITAS File System also provides interfaces for explicitly managing the layout of extents. Using a programmer interface or a command, one can choose a fixed extent size for a file, require that its blocks all be stored contiguously, and reserve space for its future growth. This allows for optimal performance for applications which manage large files, such as voice and image data. The following Buffered File System Benchmark tests are intended to show the performance advantages of extent based allocation. What the test results should indicate is that as the file I/O size increases the VERITAS File System maintains throughput by benefit of using this allocation method.

Buffered File System Benchmark

The Buffered File System Benchmark tests the standard file system performance of the VERITAS File System (VxFS) against that of the Solaris UNIX File System (UFS). These tests were run using the hardware platform mentioned in Chapter 1, along with the vxbench testing program. The reads performed in this test are standard UNIX buffer cache reads and the file system used was installed as brand new, with little or no fragmentation.

VERITAS File System

11

Performance

Buffered File System Reads

File Size KB 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 160 192 224 256 320 384 448 512 1024 2048 4096 8192 16384 32768 65536 131072 262144 523072 1048576 2097144 UFS KB/sec 82 166 334 617 1315 767 1023 1280 1286 1416 1392 1554 1728 1900 2074 1881 1859 1928 1629 2025 2408 2797 2709 3345 4118 4856 5446 8179 10991 11581 11588 11451 12101 12241 12303 12470 12364 12244 12244 VxFS KB/sec 51 104 147 299 400 630 938 1057 1217 1492 1712 1934 1734 1890 2096 2280 2545 2657 2836 3478 4024 4668 5367 4979 5654 5630 6183 9670 13688 18347 22073 24475 25463 26380 26204 26671 26896 27094 26625 UFS CPU sec 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.1 0.1 0.2 0.4 0.7 1.6 2.9 6.2 13.5 26.1 52.1 VxFS CPU sec 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.1 0.1 0.2 0.4 0.7 1.5 3 6.5 15.1 30.1 62.5

Buffered File System Writes

File Size KB 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 160 192 224 256 320 384 448 512 1024 2048 4096 8192 16384 32768 65536 131072 262144 523072 1048576 2097144 UFS KB/sec 81 162 344 650 1386 2089 2638 3101 2287 2869 3244 3659 4084 4211 4939 2960 3195 3478 3445 4480 5246 6067 6613 8016 9818 8787 8321 10197 12219 14956 15290 15729 16008 15720 15546 15920 15733 15467 15326 VxFS KB/sec 165 220 657 358 716 1036 1317 1458 2185 2262 2716 3003 3437 3974 3969 4935 4908 5395 5693 7306 9065 9603 11217 12222 15621 17642 17578 27358 31196 39098 41263 39104 44400 50220 48538 43100 38389 26642 23141 UFS CPU sec 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.1 0.1 0.2 0.4 0.8 1.6 3.5 7.2 15.7 30.9 62.4 VxFS CPU sec 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.1 0.2 0.3 0.6 1.3 2.5 5.5 11.3 24.5 52.5

These buffered tests were done using the default UFS and VxFS parameters. The hardware disk configuration used was a single SCSI controller connected to 4 SCSI drives, creating a single 8 GB RAID-0 volume.

VERITAS Software

12

File System Performance

Buffered File System Tests - Graphs

Buffered File System Read Tests

30000

25000

VxFS

20000

KB/sec

15000

10000

UFS

5000

0
1 2 4 8 16 24 32 40 48 56

64

72

80

88

96

104

112

120

128

160

192

224

320

384

448

512

1024

UFS VxFS

256

2048

4096

8192

16384

32768

131072

262144

523072

1048576 1048576

Buffered File System Write Tests

60000 System Buffer Cache Limit

50000

VxFS

40000

KB/sec

30000

20000

10000

UFS

0
1 2 4 8 16 24 32 40 48 56

64

72

80

88

96

104

112

120

128

160

224

256

320

384

UFS VxFS

192

448

512

1024

2048

4096

8192

16384

32768

131072

262144

523072

VERITAS Software

13

File System Performance

2097144

File Size KB

65536

2097144

File Size KB

65536

As the results indicate, both UFS and VxFS standard buffered read throughput begins to accelerate at the 512KB size range, however the increase for VxFS is much larger than UFS peaking at almost 26 MB/sec whereas UFS almost reaches the 12 MB/sec range. Read CPU utilization indicates a very similar curve between the two file systems. The buffered write results provide a similar picture on the right side of the table. These results show the same similar increase in throughput at the 512KB size, however the increase in VxFS throughput climbs to almost 50 MB/sec while UFS manages almost 16 MB/sec. Note that during the buffered write tests, the VERITAS File System, by default, reached its maximum limit of one half the system RAM for use as a buffer cache. Since our test platform system had 256 MB of memory, by default the VERITAS File System will limit itself to not using more than one half, or in this case 128 MB, of the installed memory for system buffers. If you were to add memory to this system, this would increase this ceiling, and increase the performance throughput measurements.

Buffered Aged File System Benchmark

This next series of buffered tests, as in the previous series, uses the buffered mode of the VERITAS File System, only. The next chapter will provide test results for the VxFS Discovered Direct I/O, and Direct I/O technologies. This second series of buffered tests involves measuring the impact that external fragmentation has on file system performance. The way in which this was accomplished was to first write a file to a new file system and then perform a read on the file. Three different block sizes were used for performing these read tests. Next multiple files were simultaneously written to a new file system, creating fragmented file allocations, and then those same files were read back in combinations while the system read throughput was measured. These buffered aged file system tests were done using the default UFS and VxFS parameters. The file size used in all iterations was 256 MB. The hardware disk configuration used was 4 SCSI controllers connected to 8 SCSI drives, creating a single 16 GB RAID-0 volume.

VERITAS Software

14

File System Performance

Buffered Aged File System Reads UFS vs. VxFS

Files Written Files Read I/O Transfer Block Size KB 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4 1 1 1 1 2 1 2 1 2 1 2 3 1 2 3 1 2 3 1 2 3 4 1 2 3 4 1 2 3 4 64 512 2048 64 64 512 512 2048 2048 64 64 64 512 512 512 2048 2048 2048 64 64 64 64 512 512 512 512 2048 2048 2048 2048 UFS Buffered KB/sec 11558 11558 11584 9783 14095 9823 14037 9821 13908 9211 11432 12787 9207 11211 12564 9228 11184 12744 8461 10411 11741 12983 8504 10391 11631 13073 8415 10195 11694 12759 VxFS Buffered KB/sec 34000 33100 33717 28524 30074 28875 30911 28905 30288 24415 28053 29224 24619 28090 29213 24438 28049 29236 22174 26163 28022 28807 22246 25969 27457 29013 22175 25914 28214 28766 UFS Buffered CPU sec 6.32 6.03 6.11 6.51 15.27 6.82 14.54 6.16 14.39 6.73 15.72 22.72 6.22 14.52 23.21 6.01 14.96 23.56 6.84 15.18 24.23 33.31 7.17 14.49 23 31.12 7.32 14.8 23.46 31.24 VxFS Buffered CPU sec 6.29 6.27 6.3 6.87 15.36 6.4 15.13 6.23 15.6 6.33 15.38 24.51 5.99 14.8 23.53 6.41 15.5 24.12 6.32 15.52 24.03 32.4 6.18 14.35 23.46 32.57 6.07 14.75 23.32 31.55

Results indicate that the fragmentation effect reduces the throughput for both UFS and VxFS. However, the results indicate that VxFS begins and maintains a higher read throughput rate and is less affected by the fragmentation. The CPU time curves are almost identical for both VxFS and UFS. (In the next chapter will illustrate the VxFS Discovered Direct I/O technology which results in much lower CPU utilization, in addition to boosting large block I/O performance.) The following graphs look at these test results:

VERITAS Software

15

File System Performance

Buffered Aged File System Tests - Graphs

Buffered Aged File System Read Tests

35000

VxFS
30000

25000

KB/sec

20000

UFS
15000

10000

5000

0
64k .5m 2m 1 1 1 1 1 1 64k 64k 1 2 2 2 .5m .5m 1 2 2 2 2m 1 2 2m 2 2 64k 64k 64k 1 2 3 3 3 3 .5m .5m .5m 1 2 3 3 3 3 2m 1 3 2m 2 3 2m 3 3 64k 64k 64k 64k 1 2 3 4 4 4 4 4 .5m .5m .5m .5m 1 2 3 4 4 4 4 4 2m 1 4 2m 2 4 2m 3 4 2m 4 4

UFS VxFS

I/O Transfer Block Files Read Files Written

Buffered Aged File System Read Tests

35

30

25

CPU Seconds

20

15

10

0
64k .5m 2m 1 1 1 1 1 1 64k 64k 1 2 2 2 .5m .5m 1 2 2 2 2m 1 2 2m 2 2 64k 64k 64k 1 2 3 3 3 3 .5m .5m .5m 1 2 3 3 3 3 2m 1 3 2m 2 3 2m 3 3 64k 64k 64k 64k 1 2 3 4 4 4 4 4 .5m .5m .5m .5m 1 2 3 4 4 4 4 4 2m 1 4 2m 2 4 2m 3 4 2m 4 4

UFS VxFS

I/O Transfer Block Files Read Files Written

VERITAS Software

16

File System Performance

Chapter

Cache Policies
The UNIX operating system supplies a standard asynchronous mode for writing to files, in which data is written through a write-back page cache in system memory, to accelerate read and write access. It also calls for a synchronous mode, which writes through the cache immediately, flushing all structures to disk. Both of these methods require data to be copied between user process buffer space to kernel buffers before being written to the disk, and copied back out when read. As mentioned previously, the VERITAS File System provides two types of cache policies which enable the File System to circumvent the standard UNIX write-back page cache in system memory. The first cache policy is a feature called Direct I/O. VERITAS implemented this feature in their file system and it provides a mechanism for bypassing the UNIX system buffer cache while retaining the on disk structure of a file system. This optimization, coupled with very large extents, allows file accesses to operate at raw-disk speed. Direct I/O may be enabled via a program interface, via a mount option, or with the 2.3 version of the VERITAS File System, Direct I/O can be invoked automatically based upon the I/O size. This feature is known as Discovered Direct I/O. The second cache policy available with the 2.3 version of the VERITAS File System is the Quick I/O for Database. While Direct I/O improves many types of large I/O performance, the single writer lock policy of the UNIX OS creates a performance bottlenecks for some types of file system writes. Database application writes are particularly affected by this. Included in the VERITAS ServerSuite Database Edition 1.0, the Quick I/O for Database bypasses the single writer lock policy in the UNIX OS by representing files to applications as character devices. This allows database applications suited to utilizing raw partitions, the ability to operate like they are using a raw partition, on a file system. The most important requirement for implementing these two VERITAS File System cache policies is that all I/O requests must meet certain alignment criteria. This criteria is usually determined by the disk device driver, the disk controller, and the system memory management hardware and software. First the file offset must be aligned on a sector boundary. Next the transfer size must be a multiple of the disk sector size. Finally depending on the underlying driver, the application buffer may need to be aligned on a sector or page boundary, and subpage length requests should not cross page boundaries. The method for guaranteeing this requirement, as well as generally improving performance for RAID level volumes, is by utilizing a technique called file system alignment.

VERITAS File System

17

Performance

File System Alignment

While RAID technology increases performance in some implementations, tuning RAID systems for proper file system alignment can increase performance for most striped RAID configurations. In most commercial implementations this involves using RAID-1, RAID-5 and RAID-0+1 configurations. The technique behind file system alignment involves setting the layout of the file system across the drive array in such a manner that the workload is distributed as equally as possible. In order to accomplish this there must be a determination as to what sections of the file system to distribute, and there must be a method for aligning these sections. This can be accomplished with most modern file systems that use data grouping techniques. The beginning of a UNIX UFS cylinder group contains the metadata blocks, and these blocks tend to be centers of disk activity. Aligning the UFS file system so that the cylinder groups begin on different drives in the array will align the file system for this method. Using this technique allows the separate drives in the array to perform the highest amount of simultaneous accesses. The way in which this can be accomplished is by the setting of the RAID stripe unit size. This is the size of disk space, on each disk, that is accessed in one pass. The combined total stripe size of all the disks is known as the RAID stripe width. Setting the stripe size to 512KB on a 3 column (disk) RAID-0 array, would result in a stripe width of 1.5MB (512x3). Since the cylinder group size in UNIX is typically 4MB, setting the stripe unit size to 512K for a 3 column array as described, would mean that the beginning of each subsequent cylinder group begins on a different drive in the array. The VERITAS File Systems cylinder groups, called allocation units (AU), do not contain similar metadata blocks at the beginning of the AU. Inodes are allocated dynamically within the data blocks of each AU. What is more important in terms of file system alignment for this file system is keeping the AUs allocated on even disk boundaries. This provides increased performance throughout the file system, as well as allow Direct I/O technologies to be utilized. What this necessitates is padding the AUs so that they begin and end on even disk boundaries.

In the 2.3 version of the VERITAS File System this is done automatically if the disks are being managed by the 2.3 version of the VERITAS Volume Manager.

Direct I/O
VERITAS Software has developed a cache policy called Direct I/O in their file system and it provides a mechanism for bypassing the UNIX system buffer cache while retaining the on disk structure of a file system. The way in which Direct I/O works involves the way the system buffer cache is handled by the UNIX OS. In the UNIX operating system, once the type independent file system, or VFS, is handed a I/O request, the type dependent file system scans the system buffer cache, and verifies whether or not the requested block is in memory. If it is not in memory the type dependent file system manages the I/O processes that eventually puts the requested block into the cache.
18

VERITAS Software

File System Performance

Since it is the type dependent file system that manages this process, the VERITAS File System uses this to bypass the UNIX system buffer cache. Once the VERITAS File System returns with the requested block, instead of copying the contents to a system buffer page, it instead copies the block into the applications buffer space. Thereby reducing the time and CPU workload imposed by the system buffer cache. In order to ensure that Direct I/O mode is always enabled safely, all Direct I/O requests must meet certain alignment criteria. This criteria is usually determined by the disk device driver, the disk controller, and the system memory management hardware and software. First the file offset must be aligned on a sector boundary. Next the transfer size must be a multiple of the disk sector size. Finally depending on the underlying driver, the application buffer may need to be aligned on a sector or page boundary, and sub-page length requests should not cross page boundaries. Direct I/O requests which do not meet these page alignment requirements, or which might conflict with mapped I/O requests to the same file, are performed as datasynchronous I/O. This optimization, coupled with very large extents, allows file accesses to operate at near raw-disk speed.

Discovered Direct I/O

Prior to the 2.3 version of the VERITAS File System invoking Direct I/O was possible via one of two ways. The first way was via a programmatic interface, an application developer could enable their application to specifically invoke Direct I/O using the VxFS system calls. The second method allowed a system administrator to specifically mount a file system whereby all I/O performed on that file system was done via Direct I/O. Since the benefit of Direct I/O is evident at larger I/O sizes, VERITAS Software implemented a feature in their 2.3 version of the VERITAS File System, which allows Direct I/O to be invoked automatically once the file system I/O reaches a specified size. This feature is known as Discovered Direct I/O and is controlled via the vxtunefs command. The discovered_direct_iosz (Discovered Direct I/O size) parameter, which defaults to 256 KB, is the parameter that controls whether or not the file system should handle an I/O transaction with either buffered or Direct I/O. Once the file system I/O is greater than 256K in size, the file system will automatically handle the I/O as Direct I/O. The next series of tests were performed in order to compare Direct I/O and Discovered Direct I/O against the performance of a UNIX raw partition. These next two series of tests were run with bottlenecks created at two specific locations. The first series of tests were run in order to induce a bottleneck at the channel controller level. These were done by utilizing a single SCSI channel connected to 4 SCSI drives. The second series of tests were run in order to induce a bottleneck at the disk level. These were done by utilizing four SCSI channels connected to 8 SCSI drives (2 each).

Controller Limited Benchmark - 4 way Stripe

The controller limited benchmarks compare the performance throughput differences between VxFS buffered, Direct I/O, Discovered Direct I/O and UNIX raw I/O. These tests were performed on the benchmark platform described in Chapter 1, limiting the hardware to a single SCSI controller, connected to 4 disks in a striped RAID-0 array. This purposely introduced a performance bottleneck at the controller level. Theoretically a Fast / Wide SCSI channel can produce 20 MB/sec of streaming data. Due to SCSI bus
VERITAS Software

19

File System Performance

technology the real limit approaches 16 MB/sec throughput. Since each SCSI disk in the test platform can produce approximately 6.5 MB/sec, with four disks per controller this creates a bottleneck at the controller level. This was done in order to illustrate the performance differences between the tested technologies in this limited throughput environment. Controller Limited Read Tests - VxFS Buffered / Discovered Direct / Direct / UNIX raw I/O
I/O Transfer Block Size KB 64 256 512 1024 2048 VxFS Buffered KB/sec 14732 14723 14732 14730 14722 Discovered Direct I/O KB/sec 14733 14687 13684 14140 14770 9392 12867 13722 14149 14680 8582 12683 13745 14124 14898 Direct I/O KB/sec raw I/O KB/sec VxFS Buffered CPU sec 6.67 6.28 6.6 6.29 5.91 Discovered Direct I/O CPU sec 6.03 6.24 0.7 0.49 0.72 1.26 0.76 0.85 0.45 0.7 1.15 0.79 0.75 0.71 0.74 Direct I/O CPU sec raw I/O CPU sec

These controller limited tests were done using the default UFS and VxFS parameters. The file size used in all iterations was 256 MB. The hardware disk configuration used was a single SCSI controller connected to 4 SCSI drives, creating a single 8 GB RAID-0 volume. Controller Limited Read Test Graphs - VxFS Buffered / Discovered Direct / Direct / UNIX raw I/O

Controller Limited - Read Results

16000

14000

12000

10000

Throughput 8000 KB/sec

6000

4000

VxFS Buffered Discovered Direct I/O Direct I/O

2000

raw I/O

0 64 256 512 I/O Transfer Block Size KB 1024 2048

VERITAS Software

20

File System Performance

These file system read results show the benefit of Discovered Direct I/O. While the I/O size remains below the discovered direct I/O size of 256K, the file system performs standard buffered I/O. Once above that size, both Discovered Direct I/O and Direct I/O throughput performance climb along the same curve of raw I/O. Note that the raw I/O final throughput of nearly 16 MB/sec is almost the maximum available throughput given the controller limited testing. This was done in order to illustrate that in terms of its scalability the VERITAS File System can provide throughput that is very close to the actual hardware limits. This model of providing the highest realized throughput for each installed system, scales as you install the VERITAS File System on larger and faster platforms. The second interesting result is the fact that the CPU utilization is high while using standard buffered I/O in the Discovered Direct I/O category, and then drops appreciably once Direct I/O is invoked once past the 256 KB block size. This demonstrates the potential for tremendous scalability which will be realized in testing later in this chapter.

Controller Limited - Read Results

7

VxFS Buffered Discovered Direct I/O Direct I/O

4 CPU Seconds 3

raw I/O

0 64 256 512 I/O Transfer Block Size KB 1024 2048

VERITAS Software

21

File System Performance

Disk Limited Benchmark - 8 way Stripe

The first series of disk limited benchmarks compare the performance throughput differences between Solaris UNIX raw I/O with the VERITAS File System Buffered I/O, Discovered Direct I/O and Direct I/O. These tests were performed on the benchmark platform described in Chapter 1, limiting the hardware to four SCSI controllers, each connected to 2 disks in a striped RAID-0 array. This purposely introduced a performance bottleneck at the disk level. Theoretically a Fast / Wide SCSI channel can produce 20 MB/sec of streaming data. Due to SCSI bus technology the real limit approaches 16 MB/sec throughput. Since each SCSI disk in the test platform can produce approximately 6.5 MB/sec, with two disks per controller this creates a bottleneck at the disk level. Disk Limited Read Tests - VxFS Buffered / Discovered Direct I/O / Direct I/O / raw I/O
I/O Transfer Block Size KB 64 256 512 1024 2048 VxFS Buffered KB/sec 35896 36091 38342 38530 34265 Discovered Direct I/O KB/sec 34935 35799 30868 45136 47851 10595 20238 33996 45893 44927 8490 20217 29491 45686 49444 Direct I/O KB/sec raw I/O KB/sec VxFS Buffered CPU sec 5.85 5.72 5.76 5.77 5.71 Discovered Direct I/O CPU sec 5.91 5.67 0.73 0.65 0.7 1.3 0.69 0.68 0.64 0.73 1.13 0.7 0.74 0.66 0.62 Direct I/O CPU sec raw I/O CPU sec

These set of disk limited tests were done using the default UFS and VxFS parameters. The file size used in all iterations was 256 MB. The hardware disk configuration used was 4 SCSI controllers connected to 8 SCSI drives, creating a single 16 GB RAID-0 volume. This set of tests compares the file system read performance throughputs of VxFS Buffered, Discovered Direct I/O, Direct I/O and Solaris UNIX raw I/O. Here again we see the combined advantages of VxFS buffered I/O and Direct I/O performance. As soon as Discovered Direct I/O invokes the Direct I/O mode, the performance throughputs between the three technologies is very similar. We also see the same drop in CPU utilization once Discovered Direct I/O invokes the Direct I/O technology. Again in this series of tests, note that the raw I/O final throughput of nearly 52 MB/sec is the maximum available throughput given the disk limited testing (6.5 MB/sec times 8 drives). This again illustrates that in terms of its scalability the VERITAS File System can provide throughput that is very close to the actual hardware limits. This also illustrates that standard buffered technology reaches bottlenecks very quickly when pushed. This model of providing the highest realized throughput for each installed system, scales as you install the VERITAS File System on larger and faster platforms.

VERITAS Software

22

File System Performance

Disk Limited Read Test Graphs- VxFS Buffered / Discovered Direct I/O / Direct I/O / raw I/O

Disk Limited - Read Results

50000

45000

40000

35000 30000 Throughput KB/sec

25000

20000

15000 VxFS Buffered Discovered Direct I/O 5000 0 64 256 512 I/O Transfer Block Size KB 1024 2048 Direct I/O raw I/O

10000

Disk Limited - Read Results

6

VxFS Buffered Discovered Direct I/O Direct I/O raw I/O

CPU Seconds

0 64 256 512 I/O Transfer Block Size KB 1024 2048

VERITAS Software

23

File System Performance

The next series of disk limited benchmarks compare the performance throughput differences between Solaris UNIX raw I/O and UFS buffered I/O with the VERITAS File System Discovered Direct I/O and Direct I/O. Additionally this entire series of disk limited tests were done utilizing a specific mode of the vxbench program. This mode is defined here as a multiple application mode in which multiple threads perform their respective file I/O, to their own unique file. This is an application workload that is similar to a standard server environment. These tests were done using the default UFS and default VxFS parameters with one change. The vxtunefs parameter of max_direct_iosz was set to 2 MB. The file size used in all iterations was 256 MB. The hardware disk configuration used was 4 SCSI controllers connected to 8 SCSI drives, creating a single 16 GB RAID-0 volume. Disk Limited Multiple Application Read Tests
Number of Files I/O Transfer Block Size KB 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 64 64 64 64 64 256 256 256 256 256 512 512 512 512 512 1024 1024 1024 1024 1024 2048 2048 2048 2048 2048 UFS Buffered KB/sec 11818 14600 12992 13406 15244 11722 12895 12897 13427 15170 11771 14791 12952 13390 15228 11781 14482 12997 13461 15195 11795 14152 13209 13309 15266 VxFS Discovered Direct KB/sec 31797 28665 28855 30179 30553 34146 28275 29193 29762 30385 29738 24510 26814 29435 30705 45345 26498 28668 30262 31314 48987 27654 29916 30991 31650 10615 11727 16797 26025 27639 20276 18421 25344 28598 29404 30181 25317 27085 29477 30212 38373 26531 29143 30191 31264 47010 27128 30138 30737 31195 5.8 13.4 31 69.3 132.4 5.8 12.6 30.2 67.2 127.8 5.6 12.9 30.1 66.7 132.9 5.2 13.5 30.1 67.8 131.7 5.5 13.3 30.3 68.9 131 VxFS Direct KB/sec UFS Buffered CPU sec VxFS Discovered Direct CPU sec 6.6 14.9 31.2 64.2 133.6 5.8 13.9 29.9 63.2 130.8 0.6 1.6 3.4 6.6 13.4 0.8 1.6 3.1 7 13.5 0.6 1.4 3.4 7.3 13.8 1.3 2.8 5.5 9.8 21.4 0.8 1.7 3.6 7.2 15 0.8 1.8 3.1 6.4 12.9 0.8 1.7 3.3 6.2 13.5 0.7 1.3 3.1 6.5 13.5 VxFS Direct CPU sec

These test results offer a good example of the combination performance of VxFS buffered and Direct I/O, in the Discovered Direct I/O feature. Note that while the UFS throughput results remain relatively flat, VxFS Discovered Direct I/O provides better initial throughput performance than Direct I/O, and then provides a similarly increasing curve as Direct I/O. The CPU time measurements again indicate the CPU resource differences between buffered file system activity, and Direct I/O file system activity.
VERITAS Software

24

File System Performance

Disk Limited Multiple Application Read Test Graphs

Disk Limited Multiple Application Read Tests

50000 45000 40000 35000 30000 VxFS Discovered Direct

KB/sec

25000 20000 15000 10000 5000 0

64 1 64 2 64 4 64 8 64 16 256 256 256 256 256 1 2 4 8 16

VxFS Direct

UFS Buffered

UFS Buffered VxFS Direct VxFS Discovered Direct

512 512 512 512 512 1 2 4 8 16

I/O Transfer Block KB Number of Files

1024 1024 1024 1024 1 1024 2 4 8 16

2048 2048 2048 2048 1 2048 2 4 8 16

Disk Limited Multiple Application Read Tests

140

UFS Buffered

120

100

80

60

40

20 VxFS Discovered Direct VxFS Direct 0

1024 1024 1024 16 8 4 2048 2048 2048 2048 2048 8 16 4 2 1

64 1

64 2

64 4

64 8

64 16

256 1

256 256 256 256 16 8 4 2

512 512 512 512 512 16 8 4 2 1

1024 1024 2 1

VxFS Direct VxFS Discovered Direct UFS Buffered

I/O Transfer Block KB Number of Files

VERITAS Software

25

File System Performance

CPU Seconds

Tuning UFS File Systems

Most benchmark tests in this report were run with little changes to the default file system parameters. However there are a number of tunable settings for both UFS and VxFS that can be utilized to increase performance, for a given application workload. As mentioned previously in the section on UFS, the UFS cylinder groups can be aligned for improving performance of UFS throughput. UFS also contains some file system settings which can affect overall system throughput. The tuning information used in this series of tests is included in the Tuning UFS section following the graphs. The next series of disk limited tests were done comparing the Solaris UFS buffered I/O tuned for large block I/O, and the VERITAS File System Discovered Direct I/O using a multiple application mode test. These tests were done using the tuned UFS parameters (explained below) and the default VxFS parameters with one change. The vxtunefs parameter of max_direct_iosz was set to 2 MB. The file size used in all iterations was 256 MB. The hardware disk configuration used was 4 SCSI controllers connected to 8 SCSI drives, creating a single 16 GB RAID-0 volume. Disk Limited Multiple Application Read Tests - Tuned UFS Buffered / VxFS Discovered Direct
Number of Files I/O Transfer Block Size KB 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 1 2 4 8 16 64 64 64 64 64 256 256 256 256 256 512 512 512 512 512 1024 1024 1024 1024 1024 2048 2048 2048 2048 2048 Tuned UFS Buffered KB/sec 30074 36802 24339 22091 15856 31230 35576 24958 22106 15744 30845 32872 23949 21595 15846 31250 32212 24152 22009 15891 30548 35032 23831 21632 15746 VxFS Discovered Direct KB/sec 31797 28665 28855 30179 30553 34146 28275 29193 29762 30385 29738 24510 26814 29435 30705 45345 26498 28668 30262 31314 48987 27654 29916 30991 31650 Tuned UFS Buffered CPU sec 4.4 11.7 28.8 60.8 141.1 4.2 11.7 28.4 60 138.9 4.2 11.2 27.9 59.2 133.1 4.1 11.3 28.6 59.4 134.3 4.2 11.9 29 60 136 6.6 14.9 31.2 64.2 133.6 5.8 13.9 29.9 63.2 130.8 0.6 1.6 3.4 6.6 13.4 0.8 1.6 3.1 7 13.5 0.6 1.4 3.4 7.3 13.8 VxFS Discovered Direct CPU sec

VERITAS Software

26

File System Performance

Disk Limited Multiple App. Read Test Graphs - Tuned UFS Buffered / VxFS Discovered Direct

Disk Limited Multiple Application Read Tests

50000 45000 VxFS Discovered Direct 40000 35000 30000

KB/sec

25000 20000 15000 Tuned UFS Buffered 10000 5000 0

64 1 64 2 64 4 64 8 64 16 256 256 256 256 256 1 2 4 8 16

512 1

512 512 2 4

512 8

512 16

VxFS Discovered Direct Tuned UFS Buffered I/O Transfer Block KB Number of Files

1024 1024 1024 1024 1 1024 2 4 8 16

2048 2048 2048 2048 1 2048 2 4 8 16

Disk Limited Multiple Application Read Tests

160

Tuned UFS Buffered

140

120

80

60

40

20 VxFS Discovered Direct 0

1024 1024 1024 1024 1024 16 8 4 2 1 2048 2048 2048 4 2 1 2048 2048 16 8

64 1

64 2

64 4

64 8

64 16

256 1

256 2

256 4

256 8

256 16

512 1

512 2

512 4

512 8

512 16

VxFS Discovered Direct I/O Transfer Block KB Number of Files Tuned UFS Buffered

VERITAS Software

27

File System Performance

CPU Seconds

100

Tuned UFS demonstrates a pretty consistent performance throughput while VxFS again demonstrates higher throughput during a majority of the testing. These file system read performance numbers indicate that proper tuning of Solaris UFS can increase overall throughput while reducing CPU utilization over the standard default Solaris UFS. However VxFS with Discovered Direct I/O, still provides better throughput with reduced CPU utilization for large file I/O. Additionally, tuning UFS for larger file I/O does impact general purpose servers that continue to perform a mix of I/O. For example tuning UFS for large file I/O, with more than one application, can cause serious degradation in systems performance, due to the OS paging mechanism and excessive CPU utilization. A better alternative is VxFS with Discovered Direct I/O Tuning UFS There are a number of information resources available in the UNIX industry which describe tuning UFS for different application workloads. Among those, DBA Systems, Inc. of Melbourne, FL. has done a lot of work in the area of tuning UFS for performing large file I/O. A report completed by William Roeder of DBA Systems, Inc. outlines the basic steps involved in performing this tuning. The information presented there is repeated here only in summary form. In tuning UFS one of the most important settings is the maxcontig parameter. By using this to alter the number of contiguous file blocks that UFS allocates for a file, the UFS system will actually operate more like an extent based file system. Other settings that can be used for tuning UFS for large block I/O are fragsize, nbpi and nrpos. Using this information the UFS file system created for this large file I/O testing was done with the following command: >mkfs -F ufs -o nsect=80,ntrack=19,bsize=8192,fragsize=8192, cgsize=16,free=10,rps=90,nbpi=32768,opt=t,apc=0,gap=0,nrpos=1, maxcontig=1536 /dev/vol/rdsk/vol01 256880

Passing the GB/sec Barrier

Since the benchmark tests for the Direct I/O technology indicates that the scalability of the VERITAS File System could be very large, VERITAS Software recently teamed up with some other high performance computing companies to test this scalability. On November 15, 1996 teams from Sun Microsystems Computer Company; Fujitsu Limited (HPC Group); VERITAS Software Corporation; Maximum Strategy, Inc.; Genroco, Inc.; and Instrumental, Inc. gathered together to assemble the software and hardware necessary to provide the highest file system I/O possible with standard open-system components. The goal was to exceed 1 GB/sec file transfer I/O. The hardware testing platform was the following: Server Hardware 1 Sun UltraSPARC 6000 with 8 167Mhz UltraSPARC CPU's 1664 MB memory 36 I/O card slots Storage Interface Hardware:
VERITAS Software

28

File System Performance

36 - Genroco Ultra-Wide S-Bus SCSI host adapter cards Storage Hardware: 4 Terabytes of RAID-5 disk arrays, provided by Maximum Strategy Gen5-S XL Disk Storage Servers. Each disk server was attached via multiple UltraSCSI channels to the UltraSCSI host adapters on the ULTRA Enterprise I/O boards. Disk attachment was via the Ultra-Wide SCSI channels provided by the Genroco S-Bus to UltraSCSI host adapters The software platform consisted of: Solaris 2.5.1 VERITAS Volume Manager (VxVM 2.3) VERITAS File System (VxFS 2.3 Beta) VERITAS Software vxbench benchmark program Instrumental's Performance Manager (PerfStat) The UltraSPARC 6000 was configured with 4 UltraSPARC CPU boards, each containing 2 CPUs and 832 MB of RAM, installed in 4 of its 16 slots. The remaining 12 slots were installed with the UltraSPARC I/O cards, each containing 3 S-Bus card slots. This provided a total of 36 S-Bus card slots into which the 36 Genroco Ultra-Wide S-Bus SCSI host adapter cards were installed. Each of the Genroco SCSI adapters was attached to one of the six ports on the Maximum Strategy Gen5-S XL RAID-5 disk arrays. In this configuration each of the ports on the Gen5-S XLs appear to the VERITAS Volume Manager as a single drive, even though they actually consist of a RAID-5 array of disks. The VERITAS Volume Manager was then used to configure all 6 Gen5-S XLs, as a single RAID-0, 36 column array. Each column used a stripe width of 2 MB, presenting a full stripe size of 72 MB. With the machine configured in this manner the testing was performed by starting 30 processes using vxbench, with each process performing I/O on one sixth of a full stripe, or 12 MB. These 30 processes would perform successive 12 MB I/O operations in parallel on a single 2 GB file in the first series of tests. The first performance throughput numbers were measured using 30 Discovered Direct I/O threads performing reads on the same 2 GB file, multiple times. Using this method we demonstrated 960 MB/sec file system throughput. We used this same method to produce a multithreaded write test on a single large file with a throughput of 598 MB/sec. At this point the testers determined that the throughput numbers while impressive, were wholly constrained by the raw disk speed. This determination was reached since Solaris raw I/O generated the same performance throughput as VxFS. As a result, with all I/O slots in the UltraSPARC 6000 filled, testers felt that the 1024 MB/sec barrier could be reached, by hooking up additional drive arrays to the server. In order to accomplish this several smaller disk arrays were attached to the SCSI host adapters built into the UltraSPARC I/O cards. Next multiple file I/O operations were run in parallel, by performing a single I/O operation on the combined Gen5-S XL large array, and by performing a single I/O operation for each of the smaller drive arrays. The final performance throughput measured was 1049 MB/sec while performing file system reads on multiple files.

VERITAS Software

29

File System Performance

An interesting side note. The highest performance throughput measured on a single Genroco Ultra SCSI controller was 27 MB/sec. Once the system was configured with 36 controllers in parallel, the performance throughput only decreased to 26.67 MB/sec per controller card. This demonstrates continued impressive scalability for VxFS.

VERITAS Software

30

File System Performance

Chapter

Quick I/O for Database

File Systems and Databases
As applications increase in size and type of workloads, some application vendors find that the standard file system design creates performance bottlenecks for their specific application workload. Database vendors in particular realize that due to the nature of their application workload, a standard file system, designed for general purpose workloads, actually introduces performance bottlenecks, simply because of the design. As an answer to this, UNIX provides a method to completely remove the file system. This is commonly referred to as raw I/O mode or raw I/O. This mode removes the file system completely from between the application and storage devices. As a result, the application vendor must provide their own file system services within their application. Since performance is of critical importance to database servers, many installations have taken to using raw I/O for any database server installations.

UNIX raw I/O

There are three basic kinds of I/O in the UNIX OS. Block devices, like disks and tapes, character devices, like terminals and printers, and the socket interface used for network I/O. All of these I/O devices are insulated from the OS by device drivers. While character devices deal in streams of data traveling between applications and devices, block devices are noted for the fact that data travels between applications and devices in blocks. These block transfers are almost always buffered in the system buffer cache. Almost every block device supports a character interface, and these are typically called raw device interfaces, or raw I/O. The difference with this interface in that none of the I/O traveling to or from the device is buffered in the system buffer cache. A limitation with this type of I/O is that depending on the device driver, the information transferred must be made in a specific block size needed by the device driver and device. As a result, UNIX supports a raw I/O mode for applications wishing to handle the file system I/O themselves. This ability to bypass the UNIX system buffer cache allows applications to define and manage their own I/O buffer cache. Database applications benefit from this technology expressly for the reason that the standard UNIX system buffer cache policies operate in way that is inefficient for most database applications. The standard UNIX system buffer cache policy is to remove pages from the buffer cache on a least recently used algorithm. This type of cache management seems to provide good performance for a broad range of applications. After examining the technology,
VERITAS File System

31

Performance

database applications have found that the performance of the database will increase if the system buffer cache employs a most frequently used caching policy. Another database application bottleneck with file systems, is that the UNIX OS maintains a single writer / multiple reader access policy on each individual file block. This allows the OS to enforce a policy of system updates being guaranteed to be updated by a single user at a time. This would keep file blocks from being corrupted with multiple simultaneous writes. However database applications lock data updates at a much more granular level, sometimes going so far as to lock updates based upon fields in a database record. As a result, locking an entire file block for data contained in a single field slows down database updates. Bypassing the file system and using a raw I/O interface allows the database vendor to lock system writes in a manner most efficient for their application. Using raw I/O allows a database vendor to employ an I/O system that is optimized to provide the best performance for their application. The largest problem that using raw I/O creates is the fact that raw I/O disks do not contain a file system. Therefore the data on the disks cannot be accessed using file system based tools, such as backup programs. raw I/O Limitations The largest single category of limitations that exist with raw I/O partitions are their management. Since the application manages all file system services, then any file system services such as backup, administration, and restoring, must be done within the application. Most of these tools treat the raw partition as one large image rather than separate files. System backups and restores must be done as a whole image, and performing maintenance on any one section of the raw system, can be very time consuming. As database servers grow in size the management problems associated with raw partitions increase.

Quick I/O for Database

The second new cache policy available with the 2.3 version of the VERITAS File System is the Quick I/O for Database. While Direct I/O improves many types of large I/O performance, the single writer lock policy of the UNIX OS creates a performance bottlenecks for some types of file system writes. Database application writes are particularly affected by this. Included in the VERITAS ServerSuite Database Edition 1.0, the Quick I/O for Database bypasses the single writer lock policy in the UNIX OS by representing files to applications as character devices. This allows database applications suited to utilizing raw partitions, the ability to operate like they are using a raw partition, on a file system. This combines the manageability of file systems with the performance of raw partitions. The benefit of the Quick I/O for Database technology is demonstrated in the next series of benchmarks. It is important to understand that unlike Direct I/O, there are two limitations on implementing Quick I/O for Database. The first is the same as Direct I/O, the file system must be properly aligned. The second is that the file space which will be used by Quick I/O for Database must be pre-allocated which is typical for database applications.

Multiple Thread Random I/O Benchmarks

The following benchmark tests compare the multiple thread performance of VxFS Buffered I/O, Direct I/O and Quick I/O for Database with UNIX raw I/O throughput:
VERITAS Software

32

File System Performance

Multiple Thread Random I/O Read Tests

Threads VxFS Buffered KB/sec 1 4 16 1748 2858 3798 1877 3096 4062 Direct I/O KB/sec Quick I/O for DB KB/sec 1898 3122 4145 1812 2987 4018 raw I/O KB/sec VxFS Buffered CPU/sec 11.1 11.4 13.1 5 4.8 6.2 Direct I/O CPU/sec Quick I/O for DB CPU/sec 4.4 5.2 5.4 4.3 4.4 5.3 raw I/O CPU/sec

Multiple Thread Random I/O Write Tests

Threads VxFS Buffered KB/sec 1 4 16 1314 1430 1401 1986 2001 1961 Direct I/O KB/sec Quick I/O for DB KB/sec 1983 3001 3887 1906 2899 3797 raw I/O KB/sec VxFS Buffered CPU/sec 11.2 11.1 14.1 4.4 5 7.1 Direct I/O CPU/sec Quick I/O for DB CPU/sec 4.5 5 6 4.1 4.3 5.2 raw I/O CPU/sec

All of the multiple thread random tests were done using the default UFS and default VxFS parameters. The file size used in all iterations was 1 GB, the block size used in all iterations was 2 KB. The hardware disk configuration used was 4 SCSI controllers connected to 16 SCSI drives, creating four 8 GB RAID-0 volumes. Finally twenty 1 GB files were pre-allocated, 5 each to a single volume for performing these tests. These benchmarks illustrate that for reads, all technologies provide very similar throughput. In some cases Quick I/O for Database actually provides slightly better throughput than the raw partition. This is likely due to some of the alignment features inherent in the combination of VERITAS Volume Manager and the VERITAS File System. The only big difference in these benchmarks is in the marked decrease in CPU utilization changing from buffered I/O to non-buffered I/O. Again Direct I/O and Quick I/O for Database perform on par with raw I/O. However when switching to the write benchmark tests it becomes apparent how much of a performance cost the single writer lock policy in the UNIX OS incurs. It is important to note that these bottlenecks exist utilizing this type of I/O stream, due to the fact that I/O will queue up behind the UNIX locking. Note that while buffered and Direct I/O reach bottlenecks in throughput at their respective levels, the Quick I/O for Database technology demonstrates impressive throughput while circumventing this system limitation. The following are the combination graphs of these results looking at the 16 thread tests:

VERITAS Software

33

File System Performance

Multiple Thread Random I/O Comparison Test Graphs

Multiple Thread Random I/O Tests

raw I/O

Quick I/O

Direct I/O

VxFS Buffered

500

1000

1500

2000

2500

3000

3500

4000

4500

Throughput KB/sec 16 Threads - 2 KB Writes 16 Threads - 2 KB Reads

Multiple Thread Random I/O Tests

raw I/O

Quick I/O

Direct I/O

VxFS Buffered

8 CPU Seconds

10

12

14

16

16 Threads - 2 KB Writes 16 Threads - 2 KB Reads

VERITAS Software

34

File System Performance

Chapter

Direct I/O vs. Quick I/O for Database

The final series of benchmarks focuses on the new cache policy technology for large block I/O, in the VERITAS File System. These are Direct I/O, and the Quick I/O for Database. Quick I/O for Database technology introduces impressive throughput as demonstrated in the previous multiple thread tests.

Direct I/O vs. Quick I/O for Database Benchmarks

Direct I/O vs. Quick I/O for Da tabase Multiple Thread Write Tests
Threads I/O Transfer Block Size KB 1 4 8 16 32 1 4 8 16 32 1 4 8 16 32 1 4 8 16 32 1 4 8 16 32 64 64 64 64 64 256 256 256 256 256 512 512 512 512 512 1024 1024 1024 1024 1024 2048 2048 2048 2048 2048 Direct I/O KB/sec 6108 6047 6041 6181 6531 16828 16810 16811 16653 15574 28693 24070 25357 24626 22975 33813 32956 31658 30642 29958 36313 35959 35799 35305 33701 Quick I/O for DB KB/sec 6644 19736 28764 42104 51368 15829 29284 49128 46539 43378 23847 49373 46217 40537 39549 31572 49435 46231 43517 38637 36355 45835 45854 40338 38255 Direct I/O CPU sec 0.91 2.16 2.16 2.18 2.52 0.74 1.09 0.99 1.06 0.65 0.64 1.06 0.9 0.71 0.99 0.68 0.77 0.7 0.9 0.94 0.63 0.82 0.81 0.95 1.26 Quick I/O for DB CPU sec 1.27 2.42 2.41 2.82 3.03 0.8 1.52 2.3 2.36 2.24 0.69 1.1 1.43 1.97 2.63 0.74 1.11 1.57 2.22 3.14 0.71 1.39 1.82 2.92 4.4

VERITAS File System

35

Performance

These benchmark tests continued the previous chapter testing by performing multiple thread, file system write testing, comparing Direct I/O with Quick I/O for Database for large block I/O, typical of imaging and other multimedia application workloads. All of the final comparison tests were done using the default VxFS parameters. The file size used in all iterations was 256 MB. The hardware disk configuration used was 4 SCSI controllers connected to 8 SCSI drives, creating a single 16 GB RAID-0 volume. Note that the performance curves begin to separate when multiple threads are used at the 4 thread level. Quick I/O for Database continues to demonstrate better throughput as the testing continues, while Direct I/O maintains a regular decrease in performance whenever the block size is smaller. CPU utilization curves appear very similar only until the larger thread ranges are reached. There the Quick I/O for Database technology utilizes more CPU time, not for file system activity but rather this is mostly attributed to thread re-synchronization. The following are the graphs of these results: Direct I/O vs. Quick I/O for Database Multiple Thread Writ e Test Graphs

Multiple Thread Write Tests - Direct I/O vs. Quick I/O

60000

Quick I/O

50000

40000

30000

20000 Direct I/O 10000

0
1024 1024 1024 1024 1024 16 32 8 4 1 2048 2048 2048 2048 2048 16 32 8 4 1

64 1

64 4

64 8

64 16

64 32

256 1

256 4

256 8

256 16

256 32

512 1

512 4

512 8

512 16

512 32

Direct I/O Quick I/O

I/O Transfer Block KB Number of Threads

VERITAS Software

36

File System Performance

KB/sec

Multiple Thread Write Tests - Direct I/O vs. Quick I/O

4.5

3.5

3 Quick I/O 2.5

1.5 1 0.5 0
1024 1024 1024 1024 1024 16 32 8 4 1 2048 2048 2048 2048 2048 16 32 8 4 1

Direct I/O

64 1

64 4

64 8

64 16

64 32

256 1

256 4

256 8

256 16

256 32

512 1

512 4

512 8

512 16

512 32

Direct I/O Quick I/O

I/O Transfer Block KB Number of Threads

These results demonstrate an interesting fact when it comes to Direct I/O. It is shown here that as the I/O transfer size increases, the Direct I/O throughput increases. This demonstrates that when the I/O requested gets larger the UNIX single writer lock policy affect is lessened. The concept behind this result is that when using low I/O sizes the disk subsystem is waiting for work to do based on the lock bottleneck. However as the I/O size increases, the disk subsystem has more work to do, and the bottleneck changes from the file system and balances with the disk subsystem. This concept is further tested in the last series of benchmark tests. Here instead of performing multiple thread writes to a single file, the comparison testing of Direct I/O and Quick I/O for Database involves performing multiple writes to different files, simulating the activity of multiple applications. This is typical of the workload imposed in a multiuser server environment.

VERITAS Software

37

File System Performance

CPU Seconds

Direct I/O vs. Quick I/O for Database M ultiple Application Write Tests
Files Written 1 2 4 8 12 1 2 4 8 12 1 2 4 8 12 1 2 4 8 12 1 2 4 8 12 I/O Transfer Block Size KB 64 64 64 64 64 256 256 256 256 256 512 512 512 512 512 1024 1024 1024 1024 1024 2048 2048 2048 2048 2048 Direct I/O KB/sec 6122 10803 21718 25750 27154 15826 22058 27187 29542 30014 23400 28667 26731 30202 30242 31935 37247 38730 38269 37413 36134 41792 43821 43575 42729 Quick I/O for DB KB/sec 5875 10989 21176 26027 27581 15681 22534 27289 29084 29944 23683 28878 26716 30199 30223 31335 33647 36731 38260 37454 36313 43808 43474 43398 42627 Direct I/O CPU sec 0.89 2.31 4.93 10.29 15.58 1.06 1.53 3.11 6.8 10.31 0.71 1.34 2.75 6.26 9.66 0.54 1.27 3.05 5.93 8.39 0.61 1.32 2.76 5.79 8.65 Quick I/O for DB CPU sec 1.15 2.23 5.02 10.39 15.38 1.07 1.58 3.49 6.94 10.38 0.68 1.53 2.93 5.92 9.62 0.76 1.14 2.91 5.49 8.55 0.68 1.43 2.89 5.51 8.32

This time the performance curves are almost identical. An interesting result considering the different technologies. This certainly demonstrates that for large I/O in a multiuser environment, utilizing Direct I/O technology, the UNIX single writer lock policy is less of an issue on overall system throughput. The following are the graphs of these results:

VERITAS Software

38

File System Performance

Direct I/O vs. Quick I/O for Database Multiple Application Write Test Graphs

Multiple Application Write Tests - Direct I/O vs. Quick I/O

45000

40000

35000 Quick I/O 30000

25000

Direct I/O

20000 15000 10000 5000 0

1024 1024 1024 1024 1024 12 8 4 2 1 2048 2048 2048 4 2 1 2048 2048 12 8

64 1

64 2

64 4

64 8

64 12

256 1

256 2

256 4

256 8

256 12

512 1

512 2

512 4

512 8

512 12

Direct I/O Quick I/O

I/O Transfer Block KB Number of Files

Multiple Application Write Tests - Direct I/O vs. Quick I/O

16

14

12

Quick I/O Direct I/O 8

0
1024 1024 1024 12 1024 1024 8 4 2 1 2048 2048 2048 2048 2048 12 8 4 2 1

64 1

64 2

64 4

64 8

64 12

256 1

256 2

256 4

256 8

256 12

512 1

512 2

512 4

512 8

512 12

Direct I/O Quick I/O

I/O Transfer Block KB Number of Files

VERITAS Software

39

File System Performance

CPU Seconds

10

KB/sec

VERITAS Software

40

File System Performance

Chapter

Conclusion
This performance report focused on the performance of the 2.3 version of the VERITAS File System. Included here has been a discussion of the key performance components in the VERITAS File System. This report also presented a series of benchmark tests comparing the throughput and CPU utilization of different VERITAS File System component technologies, as well as some tests comparing the VERITAS File System (VxFS) with the Solaris UNIX File System (UFS). It is clear that the VERITAS File System presents technologies for improved performance at the smaller I/O sizes with features such as extent based allocation. The VERITAS File System also presents technologies for advanced performance with large I/O using the Direct I/O technology. With the release of the 2.3 version of the VERITAS File system both buffered and Direct I/O performance can be combined in one file system with the Discovered Direct I/O feature. For database implementations the Quick I/O for Database provides throughput very close to that of database servers running on raw partitions. This report outlined the performance advantages and system scalability of the VERITAS File System. Using the VERITAS File System and a Sun UltraSPARC server, VERITAS has been able to generate file system performance throughput in excess of 1 GB/sec. In closing, VERITAS Software presents software technology that provides commercial class performance, availability, and manageability. This report presents a description of the very powerful performance component of VERITAS Software. Once the performance component is understood, it is important to realize that this performance comes not as a result of a standard UNIX file system, but comes coupled with the highly available, journaled, VERITAS File System.

VERITAS File System

41

Performance

VERITAS Software

42

File System Performance