All to all bandwidth seems low
In my test i could not get a decent bandwidth for all to all tests. Example 16 nodes on juwels with 128 tasks total
Linktest Timing Results - Iteration 1:
RESULT: Min Time: 1.85739994 ms ( 4.059 GiB/s)
RESULT: Max Time: 7.64846802 ms ( 1009.476 MiB/s)
RESULT: Avg Time: 5.53686910 ms ( 1.362 GiB/s)
Linktest All-to-All Timing Results:
RESULT: Min Time: 1.43342018 s ( 5.386 MiB/s)
RESULT: Max Time: 1.55601978 s ( 4.962 MiB/s)
RESULT: Avg Time: 1.49472171 s ( 5.165 MiB/s)
RESULT: Ping-Pong--to--All-to-All Efficiency: 1.6607%
Is all to all feature really working? Efficiency > 1 should be expected. Moreover if pingpong is only one iteration, there should be only 1 time measurement, why is there min and max? If that is a single node, then it is not comparable with the all to all timing.
Edited by Max Holicki