CSCI 510: Computer Architecture

Written Assignment 1 Solutions

1. Your colleague at a CPU manufacture company suggests that, since the yield is so poor, you might
make chips more cheaply if you placed an extra core on the die and only threw out chips on which
both cores had failed. We will solve this exercise by viewing the yield as a probability of no defects
occurring in a certain area given the defect rate of 0.6 per cm2. The die size is 200 mm2. Calculate
probabilities based on each core separately (this may not be entirely accurate, since the yield equation
is based on empirical evidence rather than a mathematical calculation relating the probabilities of
finding errors in different portions of the chip). Assume the process-complexity factor is 11.
a) (7 pts) What is the probability that a defect will occur on no more than one of the two processor
cores?
b) (7 pts) If the old chip cost $18 dollars per chip, what will be the cost of the new chip, taking into
account the new area and yield?
Answer:

a) Since we are considering individual dies, wafer yield is not a factor.

100 mm2 = 1cm2 => 200 mm2 = 2cm2
Probability of a die having no defect = die yield = p = 1 / (1 + 0.6 × 2)11 = 2.2-11≈ 0.00017
Probability that both dies have defects is (1 – p) 2.
Probability that no more than one die has defects is
q = 1 – (1 – p)2 = 1 – (1 – 0.00017)2 ≈ 0.00034
b) Probability that both dies have defects is (1 – p) 2.

Let wafer cost be W, dies per water be N, of which p⋅N have no defect.
18 = W / (p⋅N)
The amount of double-die chips made from one wafer is N/2.
The total amount of good double-die chips made from one wafer is q ⋅N/2
Let the cost per double-die chip with two dies be x.
The total amount of good double-die chips made from one wafer is
x⋅q⋅N/2 = W = 18 ⋅p⋅N
=> x = 18 ⋅ 2 ⋅ p / q = 36 × 0.00017 / 0.00034 = 18

2. One critical factor in powering a server farm is cooling. If heat is not removed from the computer
efficiently, the fans will blow hot air back onto the computer, not cold air. We will look at how different
design decisions affect the necessary cooling, and thus the price, of a system. Use the following table
for your power calculations.

Component Type Product Performance Power

Processor Intel Pentium 4 2 GHz 48.9–66 W
Hard drive DiamondMax 9 7200 rpm 7.9 W read/seek, 4.0 W idle
 DRAM Kingston X64C3AD2 1 GB 184-pin 3.7 W

a) (5 pts) A cooling door for a rack costs $4000 and dissipates 14 KW (into the room; additional cost
is required to get it out of the room). How many servers with an Intel Pentium 4 processor, 1 GB
184-pin DRAM, and a single 7200 rpm hard drive can you cool with one cooling door?
b) (5 pts) You are considering providing fault tolerance for your hard drive. RAID 1 doubles the
number of disks (see Chapter 6). Now how many systems can you place on a single rack with a
single cooler?
c) (5 pts) Typical server farms can dissipate a maximum of 200 W per square foot. Given that a server
rack requires 11 square feet (including front and back clearance), how many servers from part (a)
can be placed on a single rack, and how many cooling doors are required?

Answer:

a) The worst case (i.e. the peak power consumption) should be accommodated.
The peak power consumption per system is 66 + 7.9 + 3.7 = 77.6 W
The amount of systems that can be cooled = 14000 / 77.6 = 180
(x = the largest integer no more than x)
b) The peak power consumption per system is 66 + 2 × 7.9 + 3.7 = 85.5 W
The amount of systems that can be cooled = 14000 / 85.5 = 163
c) The heat that can be dissipated from a rack is 200 × 11 = 2200 W.
The amount of systems that can be cooled = 2200 / 77.6 = 28
Since 2200 W < 14000W, only one cooling door is required.

3. Server farms such as Google and Yahoo! provide enough compute capacity for the highest request
rate of the day. Imagine that most of the time these servers operate at only 60% capacity. Assume
further that the power does not scale linearly with the load; that is, when the servers are operating
at 60% capacity, they consume 90% of maximum power. The servers could be turned off, but they
would take too long to restart in response to more load. A new system has been proposed that allows
for a quick restart but requires 20% of the maximum power while in this “barely alive” state.
a) (5 pts) How much power savings would be achieved by turning off 40% of the servers? Assume
the rest 60% have to operate at 100% capacity to process the requests.
b) (5 pts) How much power savings would be achieved by placing 40% of the servers in the “barely
alive” state? Assume the rest 60% still operate at 60% capacity as the “barely alive” servers can
quickly restart.
c) (5 pts) How much power savings would be achieved by reducing the voltage by 20% and frequency
by 40% while keeping all servers operate at 60% capacity?
d) (5 pts) How much power savings would be achieved by placing 20% of the servers in the “barely
alive” state and 20% off? Assume the rest 60% have to operate at 80% capacity and consume 97%
of maximum power.

Answer:

a) Assume the maximum power consumption is P, and the amount of servers be N.

In the old system, the power consumption is 0.9NP.
 In the new system, the power consumption is 0.6N⋅1P = 0.6NP
Power saving = (0.9NP – 0.6NP) / 0.9NP = 1/3 ≈ 33.3%
b) In the new system, the power consumption is 0.6N⋅0.9P + 0.4N⋅0.2P = 0.62NP
Power saving = (0.9NP – 0.62NP) / 0.9NP  0.311= 31.1%
c)
1 2
new power 2 × capacitance × (voltage × 0.8) × (frequence × 0.6)
= = 0.384
old power 1 2 × frequence
2 × capacitance × voltage

The saving is 1 – 0.384 = 0.616 = 61.6%

d) In the new system, the power consumption is 0.2N⋅0.2P + 0.6N⋅0.97P = 0.622NP

Power saving = (0.9NP – 0.622NP) / 0.9NP  0.309= 30.9%

4. (7 pts) In a server farm such as that used by Amazon or eBay, a single failure does not cause the entire
system to crash. Instead, it will reduce the number of requests that can be satisfied at any one time.
If a company has 10,000 computers, each with a MTTF of 40 days, and it experiences catastrophic
failure only if 1/4 of the computers fail, what is the MTTF for the system?

Answer:

According to the way MTTF of individual computers is measured,

MTTF = total # hours all computers work / # computers failed

Let x be the average amount of time the system runs until 1/4 of all computers fail.

40 = x 10000 / (10000 ⋅ ¼)

=> x = 40 / 4  10 days

5. Your company is trying to choose between purchasing the Opteron or Itanium 2. You have analyzed
your company’s applications, and 60% of the time it will be running applications similar to wupwise,
10% of the time applications similar to ammp, and 30% of the time applications similar to apsi. Refer
to Figure 1.17 in the textbook for SPEC performance measurement.
a) (5 pts) If you were choosing just based on overall SPEC performance, which would you choose and
why?
b) (5 pts) What is the weighted average of execution time ratios for this mix of applications for the
Opteron and Itanium 2?
c) (5 pts) What is the speedup of the Opteron over the Itanium 2?

Answer:
 a) Since the SPECRatio of Opteron (20.86) is less than that of Itanium 2 (27.12), the choice would be
Itanium 2.
b) The weighted execution time ratio of Opteron to Itanium 2 is
0.6 × (51.5 / 56.1) + 0.1 × (136 / 132) + 0.3 × (150 / 231)  0.649
c) For Opteron, the weighted averaged execution time is
0.6 × 51.5 + 0.1 × 136 + 0.3 × 150 = 89.5 seconds
For Itanium 2, the weighted averaged execution time is
0.6 × 56.1 + 0.1 × 132 + 0.3 × 231 = 116.16 seconds
The speed up of Opteron over Itanium 2 is 116.16 / 89.5  1.30

6. When making changes to optimize part of a processor, it is often the case that speeding up one type
of instruction comes at the cost of slowing down something else. For example, if we put in a
complicated fast floating point unit, that takes space, and something might have to be moved farther
away from the middle to accommodate it, adding an extra cycle in delay to reach that unit. The basic
Amdahl’s law equation does not take into account this trade-off.
a) (5 pts) If the new fast floating-point unit speeds up floating-point operations by, on average, 2×,
and floating-point operations take 30% of the original program’s execution time, what is the
overall speedup (ignoring the penalty to any other instructions)?
b) (5 pts) Now assume that speeding up the floating-point unit slowed down data cache accesses,
resulting in a 1.5× slowdown (or 2/3 speedup). Data cache accesses consume 10% of the execution
time. What is the overall speedup now?
c) (5 pts) After implementing the new floating-point operations, what percentage of execution time
is spent on floating-point operations? What percentage is spent on data cache accesses?

Answer:

a) The overall speedup = 1 / (0.7 + 0.3 / 2) = 1/0.85  1.18

b) The overall speedup = 1 / (0.6 + 0.3 / 2 + 0.1 × 1.5) = 1/0.9  1.11
c) The percentage of execution time spent on floating-point operations is
(0.3 / 2) / (0.6 + 0.3 / 2 + 0.1 × 1.5)  0.167 = 16.7%
The percentage spent on data cache accesses is
(0.1 × 1.5) / (0.6 + 0.3 / 2 + 0.1 × 1.5)  0.167 = 16.7%

7. When parallelizing an application, the ideal speedup is speeding up by the number of processors. This
is limited by two things: percentage of the application that can be parallelized and the cost of
communication. Amdahl’s law takes into account the former but not the latter.
a) (5 pts) What is the speedup with N processors if 80% of the application is parallelizable, ignoring
the cost of communication?
b) (5 pts) What is the speedup with 8 processors if, for every processor added, the communication
overhead is 1% of the original execution time?
 c) (5 pts) What is the speedup with 8 processors if, for every time the number of processors is
doubled, the communication overhead is increased by 1% of the original execution time?
d) (5 pts) What is the speedup with N processors if, for every time the number of processors is
doubled, the communication overhead is increased by 1% of the original execution time?
e) (5 pts) Write the general equation that solves this question: What is the number of processors
with the highest speedup in an application in which P% of the original execution time is
parallelizable, and, for every time the number of processors is doubled, the communication is
increased by 1% of the original execution time?

Answer:

a) The speedup = 1 / ((1 – 0.8) + 0.8 / N) = 1 / (0.2 + 0.8 / N)

b) The speedup = 1 / ((1 – 0.8) + 8 × 0.01 + 0.8 / 8)  2.63
c) The number of processor is doubled three times in order to have 8 processors
The speedup = 1 / ((1 – 0.8) + 3 × 0.01 + 0.8 / 8)  3.03
d) To reach N processors, the number of doubling is log2N.
The speedup = 1 / ((1 – 0.8) + 0.01 log2N + 0.8 / N)
e) The speedup = 1 / (1 – 0.01P + 0.01 log2N + 0.01P / N)
To get the number of processors with the highest speedup, the derivative (on N) of the speedup
should be 0. Therefore, the equation is
-1 (1 – 0.01P + 0.01 log2N + 0.01P / N)-2 (0.01 / (N ln2) – 0.01P/N2) = 0
(ln2 means loge2)
For the equation to be zero, it must be that
0.01 / (N ln2) – 0.01P/N2 = 0
The equation can be solved to have N = 0.01P ⋅ ln2 / 0.01 = Pln2