libVMA performance seems limited by the number of memory descriptors

Dear all,

I am evaluating libVMA/Connectx-5 with UDP stream and have a performance issue apparently related to the number memory descriptor in use.

I have 20Gb/s throughput without VMA using a buffer of 128 lines of 8192 bytes. OK

I have 50Gb/s throughput using VMA with this same setup (128 lines of 8192 bytes) OK.

but that can not scale to 277 or more lines. Some packet are missed (checked with a psn). Why ? If I force the sender to slowdown, I get the data. The more I have descriptors, the more I have to slowdown

memory is allocated aligned using buffer = memalign( nLines x bufLen)

udp sockets is read in a loop: (SO_

for j < Lines

recvfrom(s, buffer + j x bufLen) so there are nLines different memory descriptors

thank you for helping

Hi Raphael,

Could you please open a support case for this issue ?

Send an email to support@mellanox.com

Thanks,

Samer