mem-ruby: add option to disable GPU coalescer deadlock watchdog - #4
Open
Basemism wants to merge 1 commit into
Open
mem-ruby: add option to disable GPU coalescer deadlock watchdog#4Basemism wants to merge 1 commit into
Basemism wants to merge 1 commit into
Conversation
The GPU coalescer watchdog aborts the simulation when a request stays outstanding longer than deadlock_threshold. This false-triggers in slow-memory sensitivity experiments where DRAM timings are intentionally stretched and long-lived requests are expected. Add a disable_deadlock_check parameter to GPUCoalescer (default False, so the watchdog stays on by default) and gate the wakeup() deadlock scan on it. Expose it through GPU_VIPER's --disable-watchdog flag, wired to both VIPER coalescers.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add a disable_deadlock_check parameter to GPUCoalescer and gate the deadlock scan in wakeup() on this parameter. It defaults to False, keeping current behaviours unchanged.
A --disable-watchdog option is also added to GPU_VIPER, disabling the deadlock check on both GPU VIPER coalescers. Other coalescer wakeup and request-processing behaviour remains unchanged.
This is useful when investigating pathological slowdowns in gem5. In some cases, execution may still be making forward progress but take significantly longer than the watchdog threshold. Panicking as soon as the threshold is reached might mask the underlying issue and preventing the simulation from running long enough to diagnose the source of the slowdown.
(In my case detecting the 1 GiB lookup issues)