Ok, I forgot to include this link. Don't know what version Cloudera is on w.r.t BlockCache, but they are claiming that using it during merges results in critical-section (allocation) lock causing meltdown...
https://blog.cloudera.com/blog/2016/08/resolving-lock-contention-in-apache-solr-a-performance-analysis-detective-story/ Will this hold good for the latest BlockCache version of Blur too? On Fri, Dec 2, 2016 at 6:20 PM, Ravikumar Govindarajan < [email protected]> wrote: > One thing I was wondering is, does block-cache acquire locks of any kind > during reads? > > I don't use the 'read-then-cache' construct at all, so was just thinking > if it is fine to eliminate locks (if any) on the read path > > > On Mon, Oct 24, 2016 at 7:07 PM, Aaron McCurry <[email protected]> wrote: > >> On Fri, Oct 21, 2016 at 1:41 AM, Ravikumar Govindarajan < >> [email protected]> wrote: >> >> > Our application makes use of 'write-thru-block-cache' only. During >> > search/merge-reads, we have modified block-cache code to only probe the >> > block-cache and avoid inserting to it. >> > >> > In such a usage scenario, I was thinking about introducing a >> > 'readBufferSize' (default=1KB) in CacheIndexInput. From block-cache or >> > underlying file we read only 'readBufferSize' data & adjust counters >> > accordingly when it's a short-circuit read... >> > >> > You think it could be made workable? >> > >> >> Yeah it should be. >> >> >> > >> > Another idea could be to bypass the cache directory during merges and >> read >> > > directly from the hdfsdirectory. Then perhaps you could take >> advantage >> > of >> > > the SC reads without having to deal with the cache directly. >> > >> > >> > This is what we are currently evaluating & it looks to be a safe bet >> > >> >> Ok, let me know if you have any questions. >> >> >> > >> > -- >> > Ravi >> > >> > On Fri, Oct 21, 2016 at 3:26 AM, Aaron McCurry <[email protected]> >> wrote: >> > >> > > I my experience I too have used block cache sizes in the 64KB range >> for >> > the >> > > same reasons you listed. The biggest of which was because we were >> > running >> > > upwards of 100GB caches and 1K block cache sizes are not really >> possible >> > at >> > > that size. The biggest probably with the compaction is with the .tim >> > file, >> > > the rest of the files are mostly sequential reads, but because that >> file >> > is >> > > a tree it tends to jump all over the place during compaction. I would >> > > recommend if you want to speed up compaction (merges) to allow the tim >> > > files to be put into block cache during the merge (e.i. turn quiet >> reads >> > > off for those files). This of course could flow your cache with data >> > that >> > > you are about to remove, so if you have the cache space it's the >> easiest >> > > solution. >> > > >> > > Another idea could be to bypass the cache directory during merges and >> > read >> > > directly from the hdfsdirectory. Then perhaps you could take >> advantage >> > of >> > > the SC reads without having to deal with the cache directly. >> > > >> > > Aaron >> > > >> > > On Thu, Oct 20, 2016 at 3:53 AM, Ravikumar Govindarajan < >> > > [email protected]> wrote: >> > > >> > > > We have set a fairly large cacheSize of 64KB in block-cache for >> > avoiding >> > > > too many keys, gc pressure etc... >> > > > >> > > > But CacheIndexInput tries to read 64KB of data during a cache-miss & >> > > fills >> > > > up the CacheValue. When doing short-circuit-reads, this could turn >> out >> > to >> > > > be excessive no? For a comparison, lucene uses only 1KB buffers for >> the >> > > > same.. >> > > > >> > > > Do you think this will likely affect performance of searches albeit >> in >> > a >> > > > minor way? >> > > > >> > > > -- >> > > > Ravi >> > > > >> > > >> > >> > >
