Difference (from prior major revision)
Changed: 2,3c2,3
< *[[Throughput graphs of raw TCP]]
< *[[Signature graphs of raw TCP]]
to
> [[image:Throughput graphs of raw TCP]]
> [[image:Signature graphs of raw TCP]]
Changed: 5,6c5,9
< *[[Aspera, throughput graphs of raw TCP]]
< *[[Aspera, signature graphs of raw TCP]]
to
> [[image:Aspera, throughput graphs of raw TCP]]
> [[image:Aspera, signature graphs of raw TCP]]
> Finally, note that the performance indicated by the graphs shown above is to be expected from machines that are otherwise idle. If this is not the case, the performance may be seriously affected. See for example the following graphs that were recorder whilst /aspera/ and /server/ were terribly busy doing molecular dynamics (with the program NAMD) :
> [[image:Throughput of raw TCP on busy cluster]]
> [[image:Signature of raw TCP on busy cluster]]
Raw TCP is the fastest you can get. The following graphs show measurements of the raw TCP performance (between the indicated cluster nodes) in the form of two graphs : throughput and signature. The later graph indicates not only maximum bandwidth, but also latency. These graphs also show the pronounced effect of adding a (hub) or, worse, a (switch + hub) on a connection line (see the aspera-related graphs).
Following the addition of a new 30m long cable, aspera was directly connected with the main 24-port switch of the cluster and the raw TCP tests repeated. The results as expected were virtually identical with those obtained for the other cluster nodes :
Finally, note that the performance indicated by the graphs shown above is to be expected from machines that are otherwise idle. If this is not the case, the performance may be seriously affected. See for example the following graphs that were recorder whilst aspera and server were terribly busy doing molecular dynamics (with the program NAMD) :
![[Home]](/wiki.png)