ppl-accel AT lists.cs.illinois.edu
Subject: Ppl-accel mailing list
List archive
- From: TMS <tms AT charm.cs.illinois.edu>
- To: ppl-accel AT cs.illinois.edu
- Subject: [[ppl-accel] ] TMS - New Log in Task Accel Meeting
- Date: Tue, 11 Sep 2018 12:47:54 -0500
- Authentication-results: illinois.edu; spf=none smtp.mailfrom=tms AT charm.cs.illinois.edu; dmarc=none header.from=charm.cs.illinois.edu
Ronak Buch has added a new log to Task Accel Meeting
Log Entry:
Jaemin has been working on tracing using CUPTI. For some reason it's
hanging when CUPTI is initialized. He also has results for GPUDirect
RDMA, results look mostly like what we'd expect, but they still need
to be compared against what non-GPUDirect looks like (both CPU-CPU and
CPU-CPU with GPU copies at both ends).
There are still some outstanding patches for GPU stuff on Gerrit that
needs some reviews and edits, Ronak and Jaemin will work on reviewing
and addressing these comments.
Ronak is mostly waiting on Michael as far as the GPUDirect messaging
layer is concerned. He did sketch out what needs to be done for the
tracing stuff for GPU:
Add tracing to GPU manager via CUPTI
- Time when kernel is issued
- Time when kernel begins executing
- Kernel occupancy? Memory footprint?
Add CUPTI data to ck-perf tracemodes (presumably just projections and
just when building with GPU manager (#ifdefs))
- Create new file for each GPU
Add ingestion support to Projections
- Display like a comm thread
- Edit views to selectively use/not use GPU data
- Can have multiple kernels executing simultaneously, so need to
interleave display
- [[ppl-accel] ] TMS - New Log in Task Accel Meeting, TMS, 09/11/2018
Archive powered by MHonArc 2.6.19.