Skip to content

Throttle evictions to one w2r task at a time#763

Closed
devreal wants to merge 1 commit into
ICLDisco:masterfrom
devreal:limit-evictions
Closed

Throttle evictions to one w2r task at a time#763
devreal wants to merge 1 commit into
ICLDisco:masterfrom
devreal:limit-evictions

Conversation

@devreal

@devreal devreal commented Mar 28, 2026

Copy link
Copy Markdown
Contributor

No need to go berserk and have multiple concurrent evictions going.

No need to go berserk and have multiple concurrent evictions going.

Signed-off-by: Joseph Schuchart <joseph.schuchart@stonybrook.edu>
@devreal devreal requested a review from a team as a code owner March 28, 2026 00:10

@bosilca bosilca left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A single w2r task can generate at most MAX_PARAM_COUNT transfers, which is a small number compared with the total number of times that exists on modern GPUs. If we only generate a single w2r at the time, we are likely to pay the w2r latency all the time, because MAX_PARAM_COUNT data are quickly reused.

Let's do instead a middleground. We allow up to MAX_W2R tasks to coexists (where MAX_W2R is an MCA parameter). Same logic as here, except that we allow more than one by default, but users can set MAX_W2R to 1 to get this sequential-looking data eviction.

@devreal

devreal commented May 11, 2026

Copy link
Copy Markdown
Contributor Author

Superseded by #773, closing.

@devreal devreal closed this May 11, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants