Any references on using SimPoints with RISC-V and SPECCPU2017 in gem5 ?

[alvaromorenomv]

Hi everyone,

Recently, I’ve been planning to run some SPEC CPU2017 simulations using the current RISC-V infrastructure in gem5.
My initial idea was to apply SimPoint-based analysis to reduce simulation time and target representative regions of execution.

However, after some research, I haven’t found any work that specifically applies SimPoint analysis to both RISC-V and SPEC CPU2017.
The closest I found was a presentation from the gem5 workshop, which focuses on SPEC CPU2006.

So I’m reaching out to ask if anyone in the community knows of any relevant references—papers, tutorials, or repositories—that explore SimPoints with RISC-V and SPEC2017 in gem5, or if any such work is currently in progress.

Thanks in advance!

[powerjg]

@studyztp do you have any pointers here?

[studyztp]

With single-threaded, the SimPoint analysis is not different between SPEC CPU2017 and SPEC CPU2006 but the workloads in SPEC CPU2017 are larger so it will take longer to do the analysis. I have a slide on how to do the SimPoint analysis with gem5 for the gem5 2024 bootcamp (link). If I remember correctly, valgrind might support bbv analysis for RISC-V. It is faster to use valgrind to do the SimPoint analysis.
I also don't know any reference that is specifically about SimPoints with RISC-V and SPEC CPU2017 in gem5, but I've done SimPoints with RISC-V and SPEC CPU2006, and other ISAs and sampling methodologies with SPEC CPU2017 in gem5 before. I can help with this. @alvaromorenomv

[alvaromorenomv]

Thank you so much for your response! @studyztp

I’ll definitely check out the slide you mentioned from the gem5 2024 bootcamp.
I wasn’t aware of the Valgrind option, it sounds like a helpful alternative to perform the SimPoint analysis, if bbv analysis is supported for RISC-V.
So I’ll explore that direction as well.

I also need to evaluate whether it would be possible to perform the analysis on some RISC-V clusters we have access to, and then decide how to proceed.

Would it be okay then if I reach out to you here in case I run into any methodological questions during the process?

[studyztp]

No problems! Feel free to reach out. I'm happy to help!
Sorry for the late reply. I thought I replied but just realized that I haven't.

[30ZHANGjin]

For those looking for a practical guide on the SimPoint workflow within gem5, I've found this document to be very helpful. It explains the principles of creating and restoring checkpoints, switching CPUs, and generating SimPoint checkpoints in detail:
https://github.com/lshpku/gem5-fs-handbook/blob/master/gem5_Checkpoint.md

Many of the actions described in that guide are based on the logic found in gem5/configs/common/Simulation.py, which makes it a reliable reference for understanding the underlying mechanics.

Also, I'm very interested in the Valgrind-based approach mentioned by @studyztp as a faster way to generate BBVs. If anyone has a working example, a tutorial, or any progress on generating BBVs for RISC-V using Valgrind, I would be very grateful if you could share it.

[yichenhit]

We have implemented simpoint generation for 1 to multiple copies of SpecInt2017 rate and SpecInt2006 rate based on Spike, which cost about only 1 day for 1 copy and 1 month for 32 copies. If there is any need, we can discuss it.

[30ZHANGjin]

We have implemented simpoint generation for 1 to multiple copies of SpecInt2017 rate and SpecInt2006 rate based on Spike, which cost about only 1 day for 1 copy and 1 month for 32 copies. If there is any need, we can discuss it.

Hi @yichenhit,
I have a follow-up question to clarify the workflow.

My understanding is that after a BBV file (.bb.gz) is generated using a fast tool like Spike or Valgrind, it is then processed by the SimPoint tool to produce simpoints.txt and weights.txt.

My question is can these resulting simpoints.txt and weights.txt files then be used directly by a gem5 script on a RISC-V system to perform the final step — creating the SimPoint checkpoints at the specified intervals?

In other words, I'm mostly curious if that last step works seamlessly — whether the data originating from the Spike/Valgrind profiling stage is fully compatible with the gem5 checkpointing stage.

Thank you!

[yichenhit]

We have implemented simpoint generation for 1 to multiple copies of SpecInt2017 rate and SpecInt2006 rate based on Spike, which cost about only 1 day for 1 copy and 1 month for 32 copies. If there is any need, we can discuss it.

Hi @yichenhit, I have a follow-up question to clarify the workflow.

My understanding is that after a BBV file (.bb.gz) is generated using a fast tool like Spike or Valgrind, it is then processed by the SimPoint tool to produce simpoints.txt and weights.txt.

My question is can these resulting simpoints.txt and weights.txt files then be used directly by a gem5 script on a RISC-V system to perform the final step — creating the SimPoint checkpoints at the specified intervals?

In other words, I'm mostly curious if that last step works seamlessly — whether the data originating from the Spike/Valgrind profiling stage is fully compatible with the gem5 checkpointing stage.

Thank you!

Hi @30ZHANGjin,
I understand what you mean is to use Spike or Valgrind to generate simpoint information, and then use Gem5 to generate checkpoints based on this information. Is this correct?

In this approach, the workload runs successively on two different emulators, which will result in certain errors and cause the generated checkpoint to be less accurate.

A better approach is to use the same emulator to generate both simpoint information and checkpoints, such as Spike. This requires modifying the Spike code to enable it to read simpoint information. In addition, the Spike needs to be modified to support the generation of checkpoints in Gem5 format.

[30ZHANGjin]

Okay, I will try the approach with Spike.

I was just thinking that if the Spike + gem5 method could be made to work, it would save a great deal of simulation time.

Thank you!