Infiniband testing 19-dec-2012
From GlueXWiki
Testing Infiniband in Rack Room
Setup
- Have 4 nodes, 16 cores each.
- Each node has one GBit and one Infiniband port.
- Infiniband switch rated at 20 Gbit/sec.
- Running IP over Infiniband, which decreases effective throughput.
- Run ET system and event producer on one node, run remote consumers on the others.
- Event producer continually reads from file and inserts events into ET system.
- Some consumers just get events and dump them (blaster mode), others process the events.
- Note: when using Ethernet instead of IB got about 95% bandwidth of a Gbit link (950 Mbits/sec).
- Note: did not tune any IP parameters for IP over IB.
Measurements in Blaster Mode
- Single (local) producer can insert events at 380 kHz, avg event size about 7.4 kBytes.
- producer uses 98% of one core, et system uses 8%.
- Single remote consumer receives events at 74 kHz (4.4 Gbit/s), 75% cpu usage.
- producer 25%, et 23%.
- Two consumers (different nodes) 148 kHz aggregate (8.8 Gbit/s), 75% cpu.
- producer 48%, et 42%.
- Three consumers 188 kHz (11.1 Gbit/s), 63% cpu.
- producer 60%, et 58%.
- Adding more consumers on the nodes doesn't change aggregate bandwidth.
Measurements in Processing Mode
- Single remote consumer receives events at 18 kHz, 93% cpu usage.
- producer 9%, et 8%.
- Two consumers 36 kHz, 93% cpu.
- producer 15%, et 11%.
- Three consumers 53 kHz, 93% cpu.
- producer 17%, et 17%.
- Add another consumer on one of the nodes, 58 kHz, 93% cpu.
- producer 19%, et 18%.
- Add another consumer on one of the nodes, 64 kHz, 93% cpu.
- producer 30%, et 22%.
- Add another, 68 kHz, 93% cpu.
- producer 20%, et 20%.
Observations
- Seem to saturate IB producer link at 11 Gbit/sec in blaster mode, requires three consumers.
- Cannot saturate IB producer link in processing mode despite many consumers.
- Adding consumers on a node doesn't scale rate up proportionally.
- Not sure what's going on with IB. N.B. Gbit Ethernet scaled as expected.