hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ramkrishna.S.Vasudevan" <ramkrishna.vasude...@huawei.com>
Subject RE: Decreasing write speed over time
Date Tue, 19 Jun 2012 14:22:14 GMT
And is your regionserver getting hotspotted? Like all your requests are targeted to one particular
RegionServer or to one particular region alone?





From: Giorgi Jvaridze [mailto:giorgi.jvaridze@gmail.com] 
Sent: Tuesday, June 19, 2012 7:26 PM
To: user@hbase.apache.org
Subject: Re: Decreasing write speed over time


Hi thanks for answer,


I've attached image of the chart.

Write performance is going down by the time. and never goes up.

As you can see on chart there are spikes but beside that generally write performance goes
down over time.




On Tue, Jun 19, 2012 at 3:21 PM, Anoop Sam John <anoopsj@huawei.com> wrote:

  Sorry I was not able to access your chart of put perf. Is it like the performance gradually
coming down and at some point again getting better all of a sudden?  [some spikes in graph]
It may be because of HBASE-3484

You can check the memstore size that u have given and see whether the spikes coming at the
memstore flush time and then again perf going down gradually...

From: Giorgi Jvaridze [giorgi.jvaridze@gmail.com]
Sent: Monday, June 18, 2012 7:28 PM
To: user@hbase.apache.org
Subject: Decreasing write speed over time

Hello list,

I have Hbase (CDH4) setup of a master and 4 region servers (8 core 2.0GHz,
6GB of RAM and 6 100GB disks per data node). All of them are virtual

I'm writing into a table using single thread w/ auto flush off.
Ids are increased by one and then reversed. Data is a string with average
length 150 chars.

I'm measuring time needed for writing one million rows. At first I got
quite good results, it could write about 40-50K rows in a second but after
time this speed decreased.
I have a chart showing how It has been decreased by the time and data got
bigger. At the end of chart it's about 2 billion rows inserted.


Is this expected behavior for this kind of setup or something could went
wrong and I should check it up?




  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message