Measuring CPUs

Prev: my-program-is-too-slow Next: measuring-memory

CPU performance is hard to measure, due to the following techniques:

CPU and memory clocks decoupled
Multiple memory banks
Multiple data registers
Multiple instructions per word
Instruction pipelining
Multiple execution units
Speculative execution
Multiple instruction issuing
Out-of-order execution
Cache memory
Paged virtual memory
Multithreading

Latency is hard to measure due to the following, as well as power saving. Most computers nowadays are heterogeneous, with efficiency cores and performance cores, where performance cores take more energy but have more performance, and efficiency cores take less energy but have less performance.

As well, CPU clocks can be dynamically slowed down to increase efficiency, making the idea of a “cycle” hard to measure.

Thus, to measure the latency of something you have to run it quite a few times.

The code shown in the book, when optimized, shows that there’s no latency, as the code is optimized away.

The compiler can also remove dead code, and reorder code, so you have to be careful about benchmarking. Also, the compiled assembly isn’t the end all be all, because the actual compiled code is also compiled in RISC 86 ops.

Exercises

2.1 What is the latency of add?

On a superscalar processor, since many instructions can be pipelined, I would expect the cycle count of adds done in succession to approach close to 0.1 cycles, since my computer can do many more than 10 adds in a single cycle.

2.2 Run mystery1.cc both unoptimized -O0 and -O2, explain the differences:

The unoptimized version runs the code as expected, and -O2 optimizes it away.

2.3 Uncomment the fprintf, see what changes:

If i uncomment the last fprintf, I assume that it can’t optimize away the code.

2.4 Declare incr as volatile and run again:

If declared as volatile, the optimizations stop and the code looks closer to as -O0.

2.5 Make your own copy of mystery1.cc and modify it to measure the latency of cycles in a 64-bit integer add. Write down your answer.

I assume it approaches 1 cycle since the optimizations would prevent parallelism and vectorization.

2.6 Experiment with the number of loop iterations. Explain why some values do not produce meaningful results.

Some values are too small to give a meaningful result and some values are too large, and thus would take too long on a server and could be affected by other processes.

2.7 Write down your guess for latency of integer 64 bit multiply and divide and floating point multiply and divide.

Integer Multiply: ~15 cycles Integer Divide: ~20 cycles

FP multiply: ~30 cycles? FP Divide: ~40 cycles?

2.8 Now verify your guesses:

I’m not going to use the program and I’ll look at Agner Fog’s latencies:

Integer multiply: 3-4 cycles Integer Divide: 9-17 cycles

FP multiply: 6-7 cycles FP Div: 15 cycles

2.9 Have your double precision drift into underflow and overflow ranges. Point out the observed latency.

I assume that latency turns into close to nothing because overflow/underflow can be handled with a single comparison.

Prev: my-program-is-too-slow Next: measuring-memory

Table of Contents

Measuring CPUs

Exercises