 |
 |
Hi Matevz, As posted in my previous comment, delaying the read will likely push the problem elsewhere and won't get rid of it. I think the better bet is to make the Ceph layer more intellegent (assuming the vreads can be satisfied with a minimal number of 64MB reads). So, the first thing is to statistically map the vreads to see how many pages they are actually hitting. Andy On Mon, 10 Aug 2020, Matev? Tadel wrote: > Currently, there is no way to disable pass-through (for either single or vector reads). It's definitely possible to do but it's not exactly trivial as it will be the first time we'd be introducing internal xcache delays in read request processing. Assuming this will be an important use case for xcache in the future, I'm willing to bite the bullet ... but you must promise to provide testing :) > > Andy, Brian, what do you think? > > You are right that as things stand reducing block-size and increasing pfc.ram are the only two handles available to you. > > Two sort of random thoughts: > > 1. You have disabled prefetching on the cache, right (pfc.prefetch 0) ... this won't matter much when cache is overloaded as prefetching automatically stops when ram usage goes beyond 80%. > > 2. Could you throw some ssds at the older nodes? > > > > -- > You are receiving this because you were mentioned. > Reply to this email directly or view it on GitHub: > https://github.com/xrootd/xrootd/issues/1259#issuecomment-671526765 -- You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: https://github.com/xrootd/xrootd/issues/1259#issuecomment-671529035 ######################################################################## 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