i keep the handler count as default(10 in 0.94.15) thanks for your reminding, i'll add regionserver's rpc monitor but how about the handler monitor? in other words, how to measure the busy degree of handler? jstack or else?
thanks 2014-09-04 20:27 GMT+08:00 Andrey Stepachev <[email protected]>: > Hi Mike. > > Need to know how many handler you have and how many clients. > Can it happen, that you have all you handlers busy with writes? > > > On Wed, Sep 3, 2014 at 11:30 PM, Bi,hongyu—mike <[email protected]> wrote: > > > btw, i disable the block cache for the hyperloglog table to avoid the > cache > > pollution > > > > > > 2014-09-04 14:29 GMT+08:00 Bi,hongyu—mike <[email protected]>: > > > > > Hi all, > > > > > > we store serialised hyperloglog object into hbase by use of > coprocessor, > > > and the size distribution is below: > > > Row size (bytes): > > > min = 4279.00 > > > max = 770757.00 > > > mean = 67340.24 > > > stddev = 153968.88 > > > median = 14453.00 > > > 75% <= 63178.00 > > > 95% <= 336917.20 > > > 98% <= 761028.00 > > > 99% <= 767500.36 > > > 99.9% <= 770757.00 > > > count = 827 > > > > > > the this value will be update 400 times every minute but the > regionserver > > > where this table locate responsed slowly for other table's get request. > > > > > > In my option, write mob/blob should not bother the performance of > reading > > > other table's region in the same regionserver, only put pressure to > hlog > > > fsync,flush and compaction; right? > > > > > > thanks > > > > > > ps: > > > hdfs blocksize=128MB > > > memstore size=128MB > > > max hlog=64 > > > low/high water: 0.35/0.4 > > > memstore percent:0.4 > > > > > > heap size=32GB > > > > > > > > > > > > > > > -- > Andrey. >
