Hi Christian, Actual test commands are below.
write: dd if=/dev/zero of=/dev/rb0 bs=4096 read: dd if=/dev/rb0 of=/dev/null bs=4096 I will check ceph readahead too. Best regards, On Dec 22, 2016 09:40, "Christian Balzer" <ch...@gol.com> wrote: > > Hello, > > On Thu, 22 Dec 2016 09:12:44 +0700 Lazuardi Nasution wrote: > > > Hi Christian, > > > > Thank you for your explanation. Based on your suggestion, I have put > > writeback cache-mode. But, currently the write ops is more better than > read > > ops. I mean "dd if=/dev/zero of=/dev/rb0" is more better than "dd > > if=/dev/rb0 of=/dev/null". Do you know what's wrong here? > > > > You may want to spend some time on figuring out what tools to use for > testing and also what your testing aims are. > As in, bandwidth or IOPS, most people tend to run into IOPS limitations > long before they run out of bandwidth. > > FIO gives you a much better control over what you're testing and how. > > Are the above "dd"s your actual full command lines? > So you read and wrote the entire device? > > Actual results instead of "better" would be nice, too. > > Either way, Ceph is known to be "slow" when doing sequential reads, google > "ceph readahead". > Also without specifying otherwise, your results are also skewed by the > pagecache and don't necessarily reflect actual Ceph performance. > > Christian > > > Best regards, > > > > On Dec 22, 2016 08:57, "Christian Balzer" <ch...@gol.com> wrote: > > > > > > > > Hello, > > > > > > On Thu, 22 Dec 2016 01:47:36 +0700 Lazuardi Nasution wrote: > > > > > > > Hi, > > > > > > > > I'm looking for the way of setting up read only cache tier but the > > > updated > > > > objects to the backend store must be evicted from the cache (if any > on > > > the > > > > cache) and it should be promoted to cache again on next miss read. > This > > > > will make the cache never contain stale objects. Which cache-mode > should > > > I > > > > use? Any other configuration related to this case? > > > > > > > > I have read about read-only, read-forward, read-proxy, proxy and > forward > > > > cache-modes. But, I'm not so sure about the different ans which one > is > > > > right for this case. I don't use write-back cache mode since writing > > > > operations has been accelerated by SSD journals. > > > > > > > > > > > If you read: > > > http://docs.ceph.com/docs/jewel/rados/operations/cache-tiering/ > > > > > > you'll have seen the dire warnings about read-only mode, which can > deliver > > > stale data. > > > > > > I can't think of any configuration options with read-only that wouldn't > > > (potentially) negate all performance improvements and still leave a > small > > > window of staleness. > > > > > > If you have mutable data and need consistency AND want to read from the > > > cache-tier, write-back is your only valid option. > > > > > > With careful tuning you should be able to avoid having writes go to the > > > cache for objects that aren't already in it. > > > > > > Christian > > > -- > > > Christian Balzer Network/Systems Engineer > > > ch...@gol.com Global OnLine Japan/Rakuten Communications > > > http://www.gol.com/ > > > > > > -- > Christian Balzer Network/Systems Engineer > ch...@gol.com Global OnLine Japan/Rakuten Communications > http://www.gol.com/ >
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
