Throughput Tests with FZK
These throughput tests were performed together with Bruno Hoeft.
Route, RTT and packet loss
traceroute to f01-110-107-e.gridka.de (192.108.45.152), 30 hops max, 38 byte packets
1 148.187.33.2 (148.187.33.2) 0.395 ms 0.396 ms 0.436 ms
2 148.187.32.4 (148.187.32.4) 0.144 ms 0.131 ms 0.143 ms
3 swima2.cscs.ch (148.187.20.2) 0.432 ms 0.421 ms 0.436 ms
4 swiEL2-10GE-1-4.switch.ch (130.59.37.77) 3.245 ms 3.380 ms 3.248 ms
5 swiCE3-10GE-1-3.switch.ch (130.59.37.65) 3.988 ms 3.977 ms 3.988 ms
6 swiCE2-10GE-1-4.switch.ch (130.59.36.209) 4.134 ms 4.124 ms 4.137 ms
7 switch.rt1.gen.ch.geant2.net (62.40.124.21) 4.132 ms 3.977 ms 3.988 ms
8 so-7-2-0.rt1.fra.de.geant2.net (62.40.112.22) 12.123 ms 12.253 ms 12.267 ms
9 dfn-gw.rt1.fra.de.geant2.net (62.40.124.34) 13.357 ms 12.210 ms 12.110 ms
10 xr-fzk1-te2-3.x-win.dfn.de (188.1.145.50) 15.053 ms 15.055 ms 14.922 ms
11 kr-fzk1.g-win.dfn.de (188.1.38.222) 15.203 ms 15.055 ms 15.075 ms
12 * * *
13 f01-110-107-e.gridka.de (192.108.45.152) 14.938 ms 14.842 ms 14.883 ms
--- f01-110-107-e.gridka.de ping statistics ---
43 packets transmitted, 43 received, 0% packet loss, time 42417ms
rtt min/avg/max/mdev = 14.866/15.028/15.313/0.141 ms, pipe 2
Influence of TCP window size and number of parallel streams
Bruno Hoeft from FZK found an optimal throughput of 700 Mbit/sec (with 4 parallel streams and a tcp window size of 4 MB) for sending to CSCS.
My own runs were performed starting at 2006-12-05 22:54h. There was some load from the worker nodes, but the CPU utilization was quite low.
But since
iperf
tests memory to memory, and the cluster bandwidth was not nearly used, the impact can be considered small.
Upload to FZK:
Download from FZK:
--
DerekFeichtinger - 06 Dec 2006