Hi Andy, I have repeated the proxy test with the same xrootd configuration file on some other machine and I can confirm that it is very likely the problem is in the proxy server implementation. This plot shows network traffic of the proxy server http://uaf-2.t2.ucsd.edu/~matevz/tmp/ReadFromProxy.png for jobs with vector and single reads. Since proxy server could not handle 300 vector reads I made another test -- run the same 300 jobs reading directly from origin. The performance was as expected http://uaf-2.t2.ucsd.edu/~matevz/tmp/ReadvDirectlyToOrigin.png I have also checked that the proxy machine can handle expected traffic in both directions using iperf http://uaf-2.t2.ucsd.edu/~matevz/tmp/ReadBasicTest.png Proxy server code was taken from the master branch. Thanks, Alja On 01/21/14 17:46, Alja Mrak-Tadel wrote: > Hi Andy, >> >> The way I read the graph is that when you put a bigger load on the >> proxy server it slows down. I wouldn't think that is unusual. The >> proxy machine may be overloaded at that point. It's unlikely to be the >> nic because we know that it can do at least 360mbps. So it might be >> the CPU (easy to check). Alternatively, it could be the disk drive at >> the end point that is overloaded with 300 jobs. >> > The proxy server had low CPU load. I did not check disk IO, but I guess > this is negligible because caching on the proxy was not enabled, neither > disk nor memory ... also, no prefetching of any kind. >> But let's say the above isn't the case because 300 jobs in single read >> perform better than 150 jobs in single read. That would point to the >> fact that either proxy is unrolling the vector reads or that the >> server is unrolling them. I suspect both because the readv passthrough >> is not available until Release 4 which has not been released yet. >> Sever-side unrolling, as unpleasant as it is, would not show such an >> effect but proxy-side unrolling most certainly would. >> >> The take-away: try the test again with a proxy server based on git head. > > I run proxy test from master branch, which I updated 34 days ago. I will > run the same set of jobs from the machine where the proxy was running > directly to the data servers (mostly at FNAL) to make sure the problem > is really in the proxy server. > > Thanks, > Alja > > > >> On Tue, 21 Jan 2014, Alja Mrak-Tadel wrote: >> >>> Hi, >>> >>> There was an issue at RAL T1 where they are trying to use proxy >>> server as their nodes do not have outgoing connectivity. >>> >>> Trying to reproduce this at UCSD, I have been running test on a proxy >>> server without any caching using the following configuration file: >>> http://uaf-2.t2.ucsd.edu/~alja/onlyproxy.cfg >>> >>> >>> On the plot http://uaf-2.t2.ucsd.edu/~matevz/tmp/traffic.png is >>> network traffic of the proxy server during four consecutive tests: >>> - 300 jobs reading 2.4MB every 10s in 128 chunk-vector-read >>> - 300 jobs reading 2.4MB every 10s in a single read >>> - 150 jobs reading 2.4MB every 10s in a single read >>> - 150 jobs reading 2.4MB every 10s in 128 chunk-vector-read >>> >>> >>> Is there something obvious that we missed in the configuration of >>> proxy-only server? How should we go about trying to find the >>> bottlenecks here? >>> >>> Thanks, >>> Alja >>> >>> ######################################################################## >>> Use REPLY-ALL to reply to list >>> >>> To unsubscribe from the XROOTD-DEV list, click the following link: >>> https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=XROOTD-DEV&A=1 >>> > > ######################################################################## > Use REPLY-ALL to reply to list > > To unsubscribe from the XROOTD-DEV list, click the following link: > https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=XROOTD-DEV&A=1 ######################################################################## Use REPLY-ALL to reply to list To unsubscribe from the XROOTD-DEV list, click the following link: https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=XROOTD-DEV&A=1