Thanks Liyin for sharing your use cases. Related to those, I was thinking of two improvements: - AFAIK, MySQL keeps the compressed and uncompressed versions of the blocs in its block cache, failing over the compressed one if decompressed one gets evicted. With very large heaps, maybe keeping around the compressed blocks in a secondary cache makes sense? - A compaction will trash the cache. But maybe we can track keyvalues (inside cached blocks are cached) for the files in the compaction, and mark the blocks of the resulting compacted file which contain previously cached keyvalues to be cached after the compaction. I have to research the feasibility of this approach.
Enis On Sun, Mar 24, 2013 at 10:15 PM, Liyin Tang <liyin.t...@fb.com> wrote: > Block cache is for uncompressed data while OS page contains the compressed > data. Unless the request pattern is full-table sequential scan, the block > cache is still quite useful. I think the size of the block cache should be > the amont of hot data we want to retain within a compaction cycle, which is > quite hard to estimate in some use cases. > > > Thanks a lot > Liyin > ________________________________________ > From: lars hofhansl [la...@apache.org] > Sent: Saturday, March 23, 2013 10:20 PM > To: user@hbase.apache.org > Subject: Re: Does HBase RegionServer benefit from OS Page Cache > > Interesting. > > > 2) The blocks in the block cache will be naturally invalid quickly after > the compactions. > > Should one keep the block cache small in order to increase the OS page > cache? > > Does you data suggest we should not use the block cache at all? > > > Thanks. > > -- Lars > > > > ________________________________ > From: Liyin Tang <liyintang2...@gmail.com> > To: user@hbase.apache.org > Sent: Saturday, March 23, 2013 9:44 PM > Subject: Re: Does HBase RegionServer benefit from OS Page Cache > > We (Facebook) are closely monitoring the OS page cache hit ratio in the > production environments. My experience is if your data access pattern is > very random, then the OS page cache won't help you so much even though the > data locality is very high. On the other hand, if the requests are always > against the recent data points, then the page cache hit ratio could be much > higher. > > Actually, there are lots of optimizations could be done in HDFS. For > example, we are working on fadvice away the 2nd/3rd replicated data from OS > page cache so that it potentially could improve your OS page cache by 3X. > Also, by taking advantage of the tiered-based compaction+fadvice in HDFS, > the region server could keep more hot data in OS page cache based on the > read access pattern. > > Another separate point is that we probably should NOT reply on the > memstore/block cache to keep hot data. 1) The more data in the memstore, > the more data the region server need to recovery from the server failures. > So the tradeoff is the recovery time. 2) The blocks in the block cache will > be naturally invalid quickly after the compactions. So region server > probably won't be benefit from large JVM size at all. > > Thanks a lot > Liyin > > On Sat, Mar 23, 2013 at 6:13 PM, Ted Yu <yuzhih...@gmail.com> wrote: > > > Coming up is the following enhancement which would make MSLAB even > better: > > > > HBASE-8163 MemStoreChunkPool: An improvement for JAVA GC when using MSLAB > > > > FYI > > > > On Sat, Mar 23, 2013 at 5:31 PM, Pankaj Gupta <pankaj.ro...@gmail.com > > >wrote: > > > > > Thanks a lot for the explanation. It's good to know that MSlab is > stable > > > and safe to enable (we don't have it enable right now, we're using > 0.92). > > > This would allow us to more freely allocate memory to HBase. I really > > > enjoyed the depth of explanation from both Enis and J-D. I was indeed > > > mistakenly referring to HFile as HLog, fortunately you were still able > > > understand my question. > > > > > > Thanks, > > > Pankaj > > > On Mar 21, 2013, at 1:28 PM, Enis Söztutar <enis....@gmail.com> wrote: > > > > > > > I think the page cache is not totally useless, but as long as you can > > > > control the GC, you should prefer the block cache. Some of the > reasons > > of > > > > the top of my head: > > > > - In case of a cache hit, for OS cache, you have to go through the DN > > > > layer (an RPC if ssr disabled), and do a kernel jump, and read using > > the > > > > read() libc vs for reading a block from the block cache, only the > > HBase > > > > process is involved. There is no process switch involved and no > kernel > > > > jumps. > > > > - The read access path is optimized per hfile block. FS page > boundaries > > > > and hfile block boundaries are not aligned at all. > > > > - There is very little control to the page cache to cache / not cache > > > > based on expected access patterns. For example, we can mark META > region > > > > blocks, and some column families, and hfile index blocks always > cached > > or > > > > cached with high priority. Also, for full table scans, we can > > explicitly > > > > disable block caching to not trash the current working set. With OS > > page > > > > cache, you do not have this control. > > > > > > > > Enis > > > > > > > > > > > > On Wed, Mar 20, 2013 at 10:30 AM, Jean-Daniel Cryans < > > > jdcry...@apache.org>wrote: > > > > > > > >> First, MSLAB has been enabled by default since 0.92.0 as it was > deemed > > > >> stable enough. So, unless you are on 0.90, you are already using it. > > > >> > > > >> Also, I'm not sure why you are referencing the HLog in your first > > > >> paragraph in the context of reading from disk, because the HLogs are > > > >> rarely read (only on recovery). Maybe you meant HFile? > > > >> > > > >> In any case, your email covers most arguments except for one: > > > >> checksumming. Retrieving a block from HDFS, even when using short > > > >> circuit reads to go directly to the OS instead of passing through > the > > > >> DN, will take quite a bit more time than reading directly from the > > > >> block cache. This is why even if you disable block caching on a > family > > > >> that the index and root blocks will still be block cached, as > reading > > > >> those very hot blocks from disk would take way too long. > > > >> > > > >> Regarding your main question (how does the OS buffer help?), I don't > > > >> have a good answer. It kind of depends on the amount of RAM you have > > > >> and what your workload is like. As a data point, I've been > successful > > > >> running with 24GB of heap (50% dedicated to the block cache) with a > > > >> workload consisting mainly of small writes, short scans, and a > typical > > > >> random read distribution for a website. I can't remember the last > time > > > >> I saw a full GC and it's been running for more than a year like > this. > > > >> > > > >> Hope this somehow helps, > > > >> > > > >> J-D > > > >> > > > >> On Wed, Mar 20, 2013 at 12:34 AM, Pankaj Gupta < > > pankaj.ro...@gmail.com> > > > >> wrote: > > > >>> Given that HBase has it's own cache (block cache and bloom filters) > > and > > > >> that all the table data is stored in HDFS, I'm wondering if HBase > > > benefits > > > >> from OS page cache at all. In the set up I'm using HBase Region > > Servers > > > run > > > >> on the same boxes as the HDFS data node. In such a scenario if the > > > >> underlying HLog files lives on the same machine then having a > healthy > > > >> memory surplus may mean that the data node can serve underlying file > > > from > > > >> page cache and thus improving HBase performance. Is this really the > > > case? > > > >> (I guess page cache should also help in case where HLog file lives > on > > a > > > >> different machine but in that case network I/O will probably drown > the > > > >> speedup achieved due to not hitting the disk. > > > >>> > > > >>> I'm asking because if page cache were useful then in an HBase set > up > > > not > > > >> utilizing all the memory on the machine for the region server may > not > > be > > > >> that bad. The reason one would not want to use all the memory for > > region > > > >> server would be long garbage collection pauses that large heap size > > may > > > >> induce. I understand that work has been done to fix the long pauses > > > caused > > > >> due to memory fragmentation in the old generation, mostly concurrent > > > >> garbage collector by using slab cache allocator for memstore but > that > > > >> feature is marked experimental and we're not ready to take risks > yet. > > > So if > > > >> the page cache was useful in any way on Region Servers we could go > > with > > > >> less memory for RegionServer process with the understanding that > free > > > >> memory on the machine is not completely going to waste. Thus my > > > curiosity > > > >> about utility of os page cache to performance of HBase. > > > >>> > > > >>> Thanks in Advance, > > > >>> Pankaj > > > >> > > > > > > > > >
