i worked to benchmark hbase performance this weekend as i never used to do this befoer.after completion of these tasks,i found some good test methods and implied interesting.
some conclusions:
---- app level ---
a.hbase 94.2 is little faster then 94.16 in writing;but in the opposite in the read case(so maybe thies is the cause why the network bandwidh is more clear then the later,see downgrade hbase from 0.94.16 to 0.94.2 )
b.as knowned knowledge,random write should be faster then sequantial one,but in this casees,i used only one client in more types,so its not explict;and in read types,sequantial is mutch faster then random.
so in general environment,i think hbase will be represnted like below:
random write > sequaltial write > sequentail read > random read
('>' is meant by throughput)
c.use presplit will be much faster than not when import huge data using multi-clients
d.cirtial two numbers are showed in figure.
e.the reuslt maybe affect by the host environment(hard/solft wares),jvm settings,network(if run in distributed cluster)
f.you MUST run the same conditions to compare both versions,else it is meanless.
g.using more data or more elapse time to run will be more correct in in theory.
----- host level ---
h.when running in mbp(ssd,maybe 800MB/sec at least),the io is not the bottle,so it will be much faster than common disks in linux.in case 1,it is 24 ms in mbp but 58 ms in linux,so i think the io cose in later will be think as :
58 - 24 ~ 30ms
=======
some questionable points
1.rows in the script is 1M by default,but i always see twice this number showed in GUI in reading case;
and when writing,this is about 1/5 to 1/10 percentage of 1M which is expected to write to it.
anyone can give me a way to figure out them?thanks
====== TODO hbase 96.x
again ,if you want to test hbase/hadoop... completely,u will spend lot of time to plan,run,compare and adjust parameters for optimizating of it!
ref: