Hi, I have done this at a customer site to overcome the 0.90.x slow WAL performance. With one RS per DN we bottlenecked, with 5-7 RS per DN we were able to hit the target rate.
Please note that we did this in lieu of the proper built-in options like WAL compression, multiple WAL, or n-way writes for WALs. Lars On Jun 16, 2012, at 10:53, Em <[email protected]> wrote: > Stack, > > I have no issues with HBase, the question is purely theoretical. > >> So, you intend doubling the datanode instances per machine too? > Everything else would not make sense to me, or what do you think? > > Thanks for your feedback! > > Regards, > Em > > Am 16.06.2012 07:12, schrieb Stack: >> On Fri, Jun 15, 2012 at 10:50 AM, Em <[email protected]> wrote: >>> what are your experience with having more than one HBase-instance per >>> machine? >>> >> >> My experience is ittle. >> >> Its been done though by others on this list. >> >>> One reason to do this could be that you got i.e. 32 GB RAM and want to >>> give 12-14 GB of RAM to each HBase instance so that you got little >>> issues with Garbage Collection while using the available RAM-capacity >>> for caching. >>> >> >> You should leave some RAM for the file system cache. >> >> Are you unable to tame your GCs? Is that why the two servers per node? >> >>> One thing I worry about is replication of data. Say two of three >>> replicas reside at the same box in two distinct HBase instances. >> >> So, you intend doubling the datanode instances per machine too? >> >> St.Ack >>
