50 likes | 99 Views
Intel MPI OFA Experience. Alexander.Supalov@intel.com German.Voronov@intel.com Sean Hefty (don’t shoot the messenger). Performance. OFA uDAPL is the best fit for Intel MPI uDAPL within 10% of native verbs Mellanox: 2.8 vs. 2.6 usecs Qlogic: 4.5 vs 4.1 usecs
E N D
Intel MPI OFA Experience Alexander.Supalov@intel.com German.Voronov@intel.com Sean Hefty (don’t shoot the messenger)
Performance • OFA uDAPL is the best fit for Intel MPI • uDAPL within 10% of native verbs • Mellanox: 2.8 vs. 2.6 usecs • Qlogic: 4.5 vs 4.1 usecs • The results are NetPipe uDAPL vs. verbs
Usability • uDAPL compatibility required between Linux and Windows • Calling conventions must match • Interoperability testing needed • Need to know Endianess and size of keys • uDAPL requires IPoIB • Can we do without?
Usability • Improve connection scaling • Longer timeouts • Faster connection setup • Want fewer user controls settable only by root • Is there a way to speed up distros adoption rate?
uDAPL • uDAPL provider behavior differs from vendor to vendor • Intel MPI needs special treatment for each provider • Different interpretations of the DAT standard • iWarp: active side of connection must send the data first • Want support for uDAPL 2.0 spec with 1.x compatible library