Counting CPU Clock Cycles

[MichaelLewis]

Can anyone recommend a way to count CPU cycles? Should we be using some facilities in std::chrono instead of the TSC register?

I've read about the TSC register that can be read using the RDTSC instruction. However, I see on Wikipedia and Stackoverflow that RDTSC has a lot of flaws on modern CPUs. For example this quote:

"The Time Stamp Counter was once an excellent high-resolution, low-overhead way for a program to get CPU timing information. With the advent of multi-core/hyper-threaded CPUs, systems with multiple CPUs, and hibernating operating systems, the TSC cannot be relied upon to provide accurate results"

 

[MichaelLewis]

Can anyone recommend a way to count CPU cycles? Should we be using some facilities in std::chrono instead of the TSC register?

I've read about the TSC register that can be read using the RDTSC instruction. However, I see on Wikipedia and Stackoverflow that RDTSC has a lot of flaws on modern CPUs. For example this quote:

"The Time Stamp Counter was once an excellent high-resolution, low-overhead way for a program to get CPU timing information. With the advent of multi-core/hyper-threaded CPUs, systems with multiple CPUs, and hibernating operating systems, the TSC cannot be relied upon to provide accurate results"

Bumping this thread in hopes of getting more visibility for a response.

 

[Daniel Duffy]

I don't do much of this in general apart from coarse-grained parallelism ball park stuff ...

What about Boost? just a guess?

CPU Timers - 1.80.0

www.boost.org

Plan B ... a commercial product .. doesn't grow on trees.

 

[MichaelLewis]

I don't do much of this in general apart from coarse-grained parallelism ball park stuff ...

What about Boost? just a guess?

CPU Timers - 1.80.0

www.boost.org

Plan B ... a commercial product .. doesn't grow on trees.

This could potentially work, thanks for the link. I’ll dive more into it to make sure it’s useful for my requirements.

 

[Paul Lopez]

Any chance you doing the build in linux? If you are you can use perf

Perf Wiki

perf.wiki.kernel.org

you can install with the below command:

sudo apt-get install linux-tools-common linux-tools-generic

say your executable can be called by doing
./mike_program

You can run:

perf stat ./mike_program

and this will give you something like the below:

$ perf stat ./mike_program

Performance counter stats for 'mike_program':

5,834,104 cycles # 0.733 GHz
2,973,423 instructions # 0.51 insn per cycle
634,204 cache-references
14,305 cache-misses # 2.255 % of all cache refs
10,502 branch-misses

0.007991768 seconds time elapsed

0.006552000 seconds user
0.001444000 seconds sys

If you want just the cycles and nothing else you can do:


perf stat -e cycles ./mike_program

Just a quick rundown of what the above output is:

• Cycles: The number of CPU cycles consumed by the mike_program command. This is a direct measure of CPU work.
• Instructions: The number of instructions executed.
• Insn per cycle: The average number of instructions executed for each cycle, an indicator of how efficiently the CPU is being used.
• Cache-references: The number of times the CPU accessed the cache.
• Cache-misses: The number of cache accesses that were not satisfied by the cache.
• Branch-misses: The number of times the CPU mispredicted the target of a branch instruction.

The final part of the output gives the time elapsed, user time, and system time for the command.

Hope it is useful! Our systems programmers and assembly champs at work love using perf.

 

[Daniel Duffy]

Are you using fine-grained or coarse-grained parallelism?

 

[MichaelLewis]

Are you using fine-grained or coarse-grained parallelism?

If anything, course grained. Although I am not even sure what I am doing would count for that.

Effectively, I am designing a system around what I know as an “Engine Model”. Where effectively the entire system is parallelized. Each thread (aka Engine) is responsible for processing events within a given symbol/instrument range.

So the system receives a book update from a venue and Engine 1 (aka thread 1) processes it because it was for AAPL. It receives another book update and Engine 2 (aka thread 2) processes it because it was for MSFT

 

[Daniel Duffy]

If anything, course grained. Although I am not even sure what I am doing would count for that.

Effectively, I am designing a system around what I know as an “Engine Model”. Where effectively the entire system is parallelized. Each thread (aka Engine) is responsible for processing events within a given symbol/instrument range.

So the system receives a book update from a venue and Engine 1 (aka thread 1) processes it because it was for AAPL. It receives another book update and Engine 2 (aka thread 2) processes it because it was for MSFT

Is this a compute-intense parallel program or some kind of asych event flow app, like a data feed pipeline?
So, what is the _problem_ ?

edit: is throughput more important; thread start/stop is costly.

 

[ExSan]

Ramtin Tajbakhsh on LinkedIn: #cplusplus #datastructuresandalgorithms #benchmark | 11 comments

This past weekend I was trying to benchmark a few optimization tricks in C++ and I came across the Google Benchmark library. It was pretty easy to setup and… | 11 comments on LinkedIn

www.linkedin.com

 

[Paul Lopez]

google's stuff is excellent thank you for the post!

 

[Paul Lopez]

@MichaelLewis sir how's it going?