Details
-
Bug
-
Resolution: Unresolved
-
Major
-
None
-
None
-
None
-
3
-
9223372036854775807
Description
Description written by Mike Marciniszyn
The issue grows out of the issues:
- Lustre creates a send queue size of 16193 with our default module parameters
- For example GPFS is two orders of magnitude smaller
- https://ibbugzilla.ph.intel.com/bugzilla/show_bug.cgi?id=142066#c1 contains an analysis of the sizing
- The size is based on the number of credits (default 128) and the LNET_MAX_IOV (hardcoded 256)
- The LNET_MAX_IOV is the key driver to the large size
- https://ibbugzilla.ph.intel.com/bugzilla/show_bug.cgi?id=142507#c37 contains the formula
- Lustre’s above sizing does not consider conns_per_peer
- I’m pretty sure credits are global regardless of the number of conns_per_peer and the QP size should be split evenly on the parallel QPs
- The TID code pre-allocates the flow structures for each send wqe
The key here is that the LNET_MAX_IOV sizes the QP based on the anticipated number of WQEs to consume a credit.
Modern Lustre memory region use (FMR,FRMR*) uses much fewer WQEs to consume a credit. Worse case 3 (vs. 256).
It is possible to detour the memory foot print by setting hfi1 max_qp_wrs to 4095. This will reduce the credits available to Lustre to 8, and will probably impact parallelism on the receive side.
Attachments
Issue Links
- mentioned in
-
Page Loading...