Thanks again for all of your input. I agree with your assessment -- in our cluster we avg 40ms for a 4k write. That's why we're adding the SSDs -- you just can't run a proportioned RBD service without them. I'll definitely give bcache a try in my test setup, but more reading has kinda tempered my expectations -- the rate of oopses and hangs on the bcache ML seems a bit high. And a 3.14 kernel would indeed still be a challenge on our RHEL6 boxes. Cheers, Dan September 4 2014 8:47 PM, "Robert LeBlanc" wrote: ? You should be able to use any block device in a bcache device. Right now, we are OK losing one SSD and it takes out 5 OSDs. We would rather have twice the cache. Our opinion may change in the future. We wanted to keep as much overhead as low as possible. I think we may spend the extra on heavier duty SSDs; less overhead not having a mirror, and less SSD drives allow us to put in more high capacity spindles in each host (we have a density need). We still have to test how much cache is optimal. I'm not sure how much 1% SSD will help and if 2% will make any difference, etc. That's why we need to test. ? My reasoning is that if we can buffer the writes, hopefully we can write to the spindles in a more linearly manner and give reads a better chance of getting serviced faster. My theory is that with all the caching already happening at KVM, etc that 1% writeback will be much more useful than 1% read cache because the reads at the OSD level will be cache misses anyway because they will be cold pages. ? Write behind is really our target, reads can be serviced from cache a good portion of the time, but writes have to always hit a disk (our current disk system is about 33% read and 66% writes), in the Ceph case (and in our config) three disks. When you add the latency from all the levels, network, kernel, buffers, disk, etc it adds up to a lot. If you are always doing large transfers, then it won't be too noticeable because bandwidth helps outweigh the latency. But when we are dealing with thousands of VMs all doing very small things like writing a couple of lines to a log, reading/writing some database pages, etc, the latency just kills the performance in a big way. Reads aren't too bad because only the primary OSD has to service the request, but on writes all three OSDs have to acknowledge the write. So I'm trying to get the absolute best write performance as I can. If I can reduce a 10 millisecond write to disk to 1 ms, then I've saved about 18 ms in the transaction (primary OSD write plus parallel write to secondary OSDs presumably). ? With bcache, I'd even like to get rid of the journal writes in a test case. Since it is hitting SSD and then being serialized to the backing disk anyway by bcache, it seems that the journal is just a double write penalty. Then it will be time to tackle the network latency, can't wait to get our Infiniband gear to test AccelIO with Ceph. ? I'm interested to see what you decide to do and what your results are. ? On Thu, Sep 4, 2014 at 12:12 PM, Dan Van Der Ster wrote: ? I've just been reading the bcache docs. It's a pity the mirrored writes aren't implemented yet. Do you know if you can use an md RAID1 as a cache dev? And is the graceful failover from wb to writethrough actually working without data loss? Also, write behind sure would help the filestore, since I'm pretty sure the same 4k blocks are being overwritten many times (from our RBD clients). Cheers, Dan On Sep 4, 2014 7:44 PM, Robert LeBlanc wrote: So far it was worked really well, we can raise/lower/disable/enable the cache in realtime and watch how the load and traffic changes. There has been some positive subjective results, but definitive results are still forth coming. bcache on CentOS 7 was not easy, makes me wish we were running Debian or Ubuntu. If there are enough reasons to train our admins on Debian/Ubuntu in addition to learning CentOS7 for customer facing boxes, we may move that way for Ceph and OpenStack, but I'm not sure how Red Hat purchasing Inktank will shift the development from Debian/Ubuntu, so we don't want to make any big changes until we have a better idea of what the future looks like. I think the Enterprise versions of Ceph (n-1 or n-2) will be a bit too old from where we want to be, which I'm sure will work wonderfully on Red Hat, but how will n.1, n.2 or n.3 run? ? Robert LeBlanc ? On Thu, Sep 4, 2014 at 11:22 AM, Dan Van Der Ster wrote: Hi Robert, That's actually a pretty good idea, since bcache would also accelerate the filestore flushes and leveldb. I actually wonder if an SSD-only pool would even be faster than such a setup... probably not. We're using an ancient enterprise n distro, so it will be a bit of a headache to get the right kernel, etc .. But my colleague is planning to use bcache to accelerate our hypervisors' ephemeral storage, so I guess that's a solved problem. Hmm... Thanks! Dan On Sep 4, 2014 6:42 PM, Robert LeBlanc wrote: We are still pretty early on in our testing of how to best use SSDs as well. What we are trying right now, for some of the reasons you mentioned already, is to use bcache as a cache for both journal and data. We have 10 spindles in our boxes with 2 SSDs. We created two bcaches (one for each SSD) and put five spindles behind it with the journals as just files on the spindle (because it is hot, it should stay in SSD). This should have the advantage that if the SSD fails, it could automatically fail to write-through mode (although I don't think it will help if the SSD suddenly fails). However, it seems that if any part of the journal is lost, the OSD is toast and needs to be rebuilt. Bcache was appealing to us because one SSD could front multiple backend disks and make the most efficient use of the SSD, it also has write around for large sequential writes so that cache is not evicted for large sequential writes which spindles are good at. Since we have a high read cache hit from KVM and other layers, this is primary intended to help accelerate writes more than reads (we are also more write heavy in our environment). ? So far it seems to help, but we are going to start more in-depth testing soon. One drawback is that bcache devices don't seem to like partitions, so we have created the OSDs manually instead if using ceph-deploy. ? I too am interested with other's experience with SSD and trying to cache/accelerate Ceph. I think the Caching pool in the long run will be the best option, but it can still use some performance tweaking with small reads before it will be really viable for us. ? Robert LeBlanc ? On Thu, Sep 4, 2014 at 10:21 AM, Dan Van Der Ster wrote: Dear Cephalopods, In a few weeks we will receive a batch of 200GB Intel DC S3700?s to augment our cluster, and I?d like to hear your practical experience and discuss options how best to deploy these. We?ll be able to equip each of our 24-disk OSD servers with 4 SSDs, so they will become 20 OSDs + 4 SSDs per server. Until recently I?ve been planning to use the traditional deployment: 5 journal partitions per SSD. But as SSD-day approaches, I growing less comfortable with the idea of 5 OSDs going down every time an SSD fails, so perhaps there are better options out there. Before getting into options, I?m curious about real reliability of these drives: 1) How often are DC S3700's failing in your deployments? 2) If you have SSD journals at a ratio of 1 to 4 or 5, how painful is the backfilling which results from an SSD failure? Have you considered tricks like increasing the down out interval so backfilling doesn?t happen in this case (leaving time for the SSD to be replaced)? Beyond the usually 5 partitions deployment, is anyone running a RAID1 or RAID10 for the journals? If so, are you using the raw block devices or formatting it and storing the journals as files on the SSD array(s)? Recent discussions seem to indicate that XFS is just as fast as the block dev, since these drives are so fast. Next, I wonder how people with puppet/chef/? are handling the creation/re-creation of the SSD devices. Are you just wiping and rebuilding all the dependent OSDs completely when the journal dev fails? I?m not keen on puppetizing the re-creation of journals for OSDs... We also have this crazy idea of failing over to a local journal file in case an SSD fails. In this model, when an SSD fails we?d quickly create a new journal either on another SSD or on the local OSD filesystem, then restart the OSDs before backfilling started. Thoughts? Lastly, I would also consider using 2 of the SSDs in a data pool (with the other 2 SSDs to hold 20 journals ? probably in a RAID1 to avoid backfilling 10 OSDs when an SSD fails). If the 10-1 ratio of SSDs would perform adequately, that?d give us quite a few SSDs to build a dedicated high-IOPS pool. I?d also appreciate any other suggestions/experiences which might be relevant. Thanks! Dan -- Dan van der Ster || Data & Storage Services || CERN IT Department -- _______________________________________________ ceph-users mailing list ceph-users at lists.ceph.com (mailto:ceph-users at lists.ceph.com) http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com (http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com) ? -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.ceph.com/pipermail/ceph-users-ceph.com/attachments/20140904/8c7ceba2/attachment.htm>
 |