On Mon, Jul 22, 2013 at 02:36:27PM +0200, Bernd Schubert wrote: > On 07/22/2013 02:47 AM, Dave Chinner wrote: > >On Fri, Jul 19, 2013 at 09:52:00PM +0200, Bernd Schubert wrote: > >>Hello Ric, hi all, > >> > >>On 07/12/2013 07:20 PM, Ric Wheeler wrote: > >>> > >>>If you have topics that you would like to add, wait until the > >>>instructions get posted at the link above. If you are impatient, feel > >>>free to email me directly (but probably best to drop the broad mailing > >>>lists from the reply). > >> > >>sorry, that will be a rather long introduction, the short conclusion > >>is below. > >> > >> > >>Introduction to the meta-cache issue: > >>===================================== > >>For quite a while we are redesigning our FhGFS storage layout to > >>workaround meta-cache issues of underlying file systems. However, > >>there are constraints as data and meta-data are distributed on > >>between several targets/servers. Other distributed file systems, > >>such as Lustre and (I think) cepfs should have the similar issues. > >> > >>So the main issue we have is that streaming reads/writes evict > >>meta-pages from the page-cache. I.e. this results in lots of > >>directory-block reads on creating files. So FhGFS, Lustre an (I > >>believe) cephfs are using hash-directories to store object files. > >>Access to files in these hash-directories is rather random and with > >>increasing number of files, access to hash directory-blocks/pages > >>also gets entirely random. Streaming IO easily evicts these pages, > >>which results in high latencies when users perform file > >>creates/deletes, as corresponding directory blocks have to be > >>re-read from disk again and again. > > > >Sounds like a filesystem specific problem. Different filesystems > >have different ways of caching metadata and respond differently to > >page cache pressure. > > > >For example, we changed XFS to have it's own metdata buffer cache > >reclaim mechanisms driven by a shrinker that uses prioritised cache > >reclaim to ensure we reclaim less important metadata buffers before > >ones that are more frequently hit (e.g. to reclaim tree leaves > >before nodes and roots). This was done because the page cache based > >reclaim of metadata was completely inadequate (i.e. mostly random!) > >and would frequently reclaim the wrong thing and cause performance > >under memory pressure to tank.... > > Well, especially with XFS I see reads all the time and btrace tells > me these are meta-reads. So far I didn't find a way to make XFS to > cache meta data permanenly and so far I didn't track that down any > further. Sure. That's what *I* want to confirm - what sort of metadata is being read. And what I see is the inode and dentry caches getting trashed, and that results in directory reads to repopulate the dentry cache.... > For reference and without full bonnie output, with XFS I got about > 800 to 1000 creates/s. > Somewhat that seems to confirm my idea not to let file systems try > to handle it themselves, but to introduce a generic way to cache > meta data. We already have generic metadata caches - the inode and dentry caches. The reason some filesystems have their own caches is that the generic caches are not always suited to the physical metadata structure of the filesystem, and hence they have their own multi-level caches and reclaim implementations that are more optimal than the generic cache mechanisms. IOWs, there isn't an "optimal" generic metadata caching mechanism that can be implemented. > >>Entirely cached hash directories (16384), which are populated with > >>about 16 million files, so 1000 files per hash-dir. > >> > >>>Version 1.96 ------Sequential Create------ --------Random Create-------- > >>>fslab3 -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete-- > >>>files:max:min /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP > >>> 60:32:32 1702 14 2025 12 1332 4 1873 16 2047 13 1266 3 > >>>Latency 3874ms 6645ms 8659ms 505ms 7257ms 9627ms > >>>1.96,1.96,fslab3,1,1374655110,,,,,,,,,,,,,,60,32,32,,,1702,14,2025,12,1332,4,1873,16,2047,13,1266,3,,,,,,,3874ms,6645ms,8659ms,505ms,7257ms,9627ms > > > >Command line parameters, details of storage, the scripts you are > >running, etc please. RAM as well, as 16 million files are going to > >require at least 20GB RAM to fully cache... > > 16 million files are only laying around in the hash directories and > are not touched at all when new files are created. So I don't know > where you take 20GB from. Each inode in memory requires between 1-1.4k of memory depending on the filesystem they belong to. Then there's another ~200 bytes per dentry per inode, and if the names are long enough, then another 64+ bytes for the name of the file held by the dentry. So caching 16 million inodes (directory or files) requires 15-25GB of RAM to cache. FWIW, have you tried experimenting with /proc/sys/vm/vfs_cache_pressure to change the ratio of metadata to page cache reclaim? You might find that all you need to do is change this ratio and your problem is solved..... > Our file names have a typical size of 21 bytes, so with a classical > ext2 layout that gives 29 bytes, with alignment that makes 32 bytes > per directory entry. Ignoring '.' and '..' we need 125000 x 4kiB > directory blocks, so about 500MB + some overhead. If the dentry cache stays populated, then how the filesystem lays out dirents on disk is irrelevant - you won't ever be reading them more than once.... > >Numbers without context or with "handwavy context" are meaningless > >for the purpose of analysis and understanding. > > I just wanted to show here, that creating new files introduces reads > when meta-data have been evicted from the cache and how easily that > can happen. From my point of view the hardware does not matter much > for that purpose. In my experience, hardware always matters when you are asking someone else to understand and reproduce your performance problem. It's often the single most critical aspect that we need to understand.... > This was with rotating disks as typically used to store huge amounts > of HPC data. With SSDs the effect would have been smaller, but even > SSDs are not as fast as in-memory-cache lookups. .... > Our customer systems usually have >=64GiB RAM and often _less_ than > 16 million files per server. But still meta-reads impact latency and > streaming performance. ..... > Please not that bonnie++ is not ideally suitable for > meta-benchmarks, but as I said above, I just wanted to demonstrate > cache evictions. Sure. On the other hand, you're describing a well known workload and memory pressure eviction pattern that can be entirely prevented from userspace. Do you reuse any of the data that is streamed to disk before it is evicted from memory by other streaming data? I suspect the data cache hit rate for the workloads you are describing (HPC and bulk data storage) is around 0%. If so, why aren't you making use of fadvise(DONTNEED) to tell the kernel it doesn't need to cache that data that is being read/written? That will prevent streaming Io from creating memory pressure, and that will prevent the hot data and metadata caches from being trashed by cold streaming IO. I know of several large scale/distributed storage server implementations that do exactly this... Remember: not all IO problems need to be solved by changing kernel code ;) Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
