Monthly gem5 Developers' Meeting - September 12th at 4pm (UTC) #1503

ivanaamit · 2024-08-22T18:17:58Z

ivanaamit
Aug 22, 2024

The monthly gem5 Developers' Meeting will be held on September 12th at 4pm UTC.

erin-le · 2024-09-12T17:10:33Z

Notes from today's meeting:

participants
- erin, jason, ivana, mahyar, harshil, giacomo, jonathan haley, hemant bhandari, daniel carvalho, richard cooper, rit
issues and prs are growing
reason for dev meetings: communicate with community about what dev team is doing
agenda: deprecating things, adding demo board, Mahyar's findings

stdlib: reproducible, accurate
eventually: want "known good configurations" - some % of accuracy against actual hardware, difficult
demo boards - establish infrastructure of accuracy testing. Not configurable/extensible, want them to be fixed.
want to show that the workloads/suites run

(showed screenshot from gem5 resources website)
want a new tab on the website that shows what the status is on the corresponding demo board
time, simtime, instr, other stats
Other cool ideas
- discussion 754
- dashboard for currently running wklds
- see what has/hasn't been tested, see if the results are reasonable
Giacomo has questions
- 1. run daily tests on board?
    a. Run only on stable - once per release cycle
- 1. How to update? - prev answer answered this
- 1. Adding fixed model: consider microarch is tightly coupled with architecture
  - divergence/convergence to real hardware when change
- 1. depends on kernel config - maybe provide config file
  - extract kernel from disk img
  - how to extract kernel from arm ubuntu? already know how to do x86
- Jason's response to 3 and 4: need to enumerate the caveats to the data
- Giacomo: make it explicit that we will not match real hardware; don't overpromise
Jason wants to work with ARM people to figure out configurations
- "configurations" covers everything
- especially kernel configs, arm platform configs, release config, hardware
- Giacomo suggests: have piece of hardware, enable all supported extensions, model based on that
  - they can't share microarch
- Jason: wants outside reviewers to increase community truse

encourage users to write python code, not bash scripts
presenting every possible parameter is bad for research
- options.py is misleading, people will continue using it if not deprecated
blockers: arm fs tests, gpu scripts
Gia: Arm fs tests don't need command line params
- some arm files in common are useful, move them? Not tightly coupled with options.py
Jason: update arm fs tests to not have params, move cores directory out of common, leave starter fs and starter se alone for now, convert to stdlib in long term
Giacomo tried to combine starter fs and stdlib, couldn't do it
- he suspects that similar issues will occur with starter se
- Jason thinks it would be possible to do starter se
GPU scripts difficult to port, discussion skipped since GPU people weren't present
Giacomo: grep file for options, extract the ones used from options.py, put in a separate config script to decouple from options.py

Arm paired memory ops
Working to develop known good config for Arm Ampere Altra
paired memory instrs cause discrepancy
short samples of execution of NPB workloads, run on both gem5 and hardware
samples:
- gem5 is from 1.6 times faster to 6.6 times slower
- gem5 accurate for data cache access
- 1x to 4x in store instr
- error in num cycles correlated to num data cache accesses
- error in data cache accesses is mostly explained by error in num load and store
giacomo comment: gem5 microop vs macroop?
translate paired mem instrs to one microop instead of possibly multiple, results in hypothetical graph
Conclusion
- error in data cache accesses is correlated and explainable by error in num load and store instrs
- we think the root cause is how PairMemOps are translated to microops in gem5
Send Giacomo slides, binaries, things needed to test

0 replies