Observations on poor transfer performance - especially with writes to Echo - have been reported.

This is a brief summary of various performance studies.

Iperf3 testing

Tested with two versions of perf.

Overall the Multi-threaded (mt) gave better throughputs, particularly with the ingress to RAL rates.

PREF=/home/tier1/jwalder/iperf-3.15-mt-beta1; LD_PRELOAD=${PREF}/lib/libiperf.so.0.0.0 ${PREF}/bin/iperf3 -c ds1-london.perf.ja.net -p 5202 -P16  --get-server-output  -4

PREF=/home/tier1/jwalder/iperf/iperf3_15; LD_PRELOAD=${PREF}/libiperf.so.0.0.0 ${PREF}/iperf3 -c ds1-london.perf.ja.net -p 5202 -P16  --get-server-output  -4

Testing and results

Single-shot tests for each result; no significant difference observed between IPv4 and IPv6
Only the results from the MT tests are shown. Testing from JISC to RAL uses the --reverse flag. i.e. the JISC end is always acting as the server

Plots below show the performance of each host from RAL to JISC and for JISC to RAL. No significant difference is seen between IPV4 and IPV6.
Between the hosts that are in production, and those not in production, little difference is also seen. For GW8, which has a bonded 50Gb/s link, there is better throughput at least for the egress speed.

For ingress tests, there appears to be a clear difference between hosts on the new network, compared to those on the legacy network. There is also a drop in performance in Ingress speeds compared to the egress rates.

The following plot combines the information from above, and shows the performance of hosts on the Legacy and New network, for Ingress and Egress rates from RAL.
This summarises both the relative performance difference between ingress and egress, as well as the difference between new and legacy network hosts