 |
 |
Hi Andy, On 10/07/11 12:22, Andrew Hanushevsky wrote: > Hi Brian, > > This should be posted as a bug. There are a number of issues: > > 1) Some modes of sending data are not counted in the file statistics, notably > when sendfile() is used. But it should be possible to add it to both counters (global and trace), right? > 2) Readv requests are not included in the detailed monitoring record. > > The latter one raises a couple of questions: > > a) Does each read segment create a monitor record entry? This could cause > excessive number of entries as readv vectors can get quite long. On the other > hand, the alternate read mechanism would simply issue multiple reads and all of > these would have been put in the monitor record. So, it may be moot. Yeah, that's too much. > b) You mention a separate entry for readv. Is this a tag on the entry? A > separate code? We can do this since readv requests were never logged in the > first place so no compatability problems. Yes, Brian and I think this would be the best option, adding a new trace type for vector reads, sending total number of bytes and number of chunks sent. Is there such a thing as vector write? > c) Should readv capture be configurable? Doesn't seem like that is useful. > > Once you post this as a bug, I will add his to your post. Are we also hit by case 1)? Should I open a new bug for this cases, too? Cheers, Matevz > Andy > > On Fri, 7 Oct 2011, Brian Bockelman wrote: > >> Hi folks, >> >> It appears there are quite a few missing bytes in the detailed xrootd >> statistics. When comparing the number of bytes sent in the sum of all >> read/write request entries with the total number of bytes in the close request >> entries, things are off by a factor of 100 (or more) for CMS jobs. >> >> Perhaps not coincidentally, vector read calls are over 99% of the bytes moved >> by CMS. Is it possible that they are not getting accounted for? >> >> Brian >> >> PS - for what it's worth, it would be nice if ReadV's could be monitored >> separately; this allows me to monitor for badly behaving jobs. >>