Data Transfer performance results
Initial performance was:
- OK/Poor from Canberra Server to master.SHAO (about 100Mbps single stream)
- Poor from Pawsey MWA FE nodes to Canberra serve (94 Mbps single stream)
- Not working from MWA FE servers to master.shao (no ping or ssh or iperf)
- OK from ASKAP server to master.SHAO (evidence of 500 Mbps ngastransfer performance)
- increased kernel TCP parameters on master.SHAO
- Parallel transfers... -P50 achieved 1.55 Gbps
- FE1, FE2 & FE3 nodes had TCP kernel paramters increased to 64 MBytes
- single-flow iperf to Canberra server gave 4.6 Gbps over 30 seconds.. up to 5.95Gbps
- Debugging took a while... connectivity OK to SHAO gateway address
- Conclusion: a problem with firewalling on master.SHAO
- Firewalling changed to white-list FE1, FE2 and FE3 nodes (plus a test server at MRO)
- single stream performance of 1.95 Gbps averae over 10 minutes with RTT of 271ms
- performance peaking at around 2.4 Gbps single stream
- performed parallel iperf3 tests
- best seems to be -P12 (12 parallel streams) - 2.73Gbps from FE2 to master-SHAO
- -P12 from ASKAP NGAS host to master.SHAO achieved 1.04 Gbps
- -P12 from FE1 to master.SHAO achieved 2.91Gbps
- -P12 from ASKAP host improved to 2.66 Gbps after TCP kernel tuning at Pawsey end
- Single stream to master.SHAO achieves about 271 Mbps
- 20 parallel transfers (-t 20) achieved 2.08 Gbps transfer over 1 minute
- Repeated Thursday morning 1.8 Gbps over 7m20s transfer, 1.7Gbps over 1 minute (-t 20)
- ASKAP reported ~ 500Mbps throughput towards SHAO before the workshop
- MWA NGAS transferrs started working on Thursday afternoon. about 1.209bps
- 108833648 kBytes * 8 bits / 720 sec = 1209262 kbps
- Future work: see whether ASKAP NGAS transfers have improved speed after the kernel tuning
- Check whether ASKAP kernel tuning has resulted in improved NGAS transfer speeds
- AARNet graph of Pawsey interface
- SHAU Graphs of upstream interface utilisation 24 hours to Thursday morning
