Local DRAM accesses
from pcm-numa does not match memory throughput from pcm-memory
#687
Unanswered
QiongwenXu
asked this question in
Q&A
Replies: 1 comment
-
there are a few more things to consider: pcm-numa measures accesses. Each read access can trigger 64 byte transfer (cache line) or up to two 64 byte transfers (read-for-ownership + write-back) for a write access. This depends on the architecture. 0.527*64 = 33 Gbyte/sec which is close to your read bandwidth measured by pcm-memory. Some of these accesses are writes and generate the additional write bandwidth (14 Gbyte/sec in pcm-memory). Hardware prefetches can also generate additional traffic. pcm-numa is not intended to measure exact memory bandwidth. It is more to assess remote/local access distribution. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi I am using PCM to measure memory channel usage. I ran a few STREAM applications (memory read/write intensive) on cpu cores 16-23. I didn't run any other applications. Then I use pcm-numa and pcm-memory to measure the memory channel usage. But
Local DRAM accesses
from pcm-numa (527 MB, interval is 1 second, hence throughput is 527 MB/s) does not match memory throughput from pcm-memory (ie, 49375.64 MB/s). Is my understanding incorrect or do you happen to know why this happens? Thanks!pcm-memory:
pcm-numa:
Beta Was this translation helpful? Give feedback.
All reactions