hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <tdunn...@maprtech.com>
Subject Re: CSLM performance Was: SILT - nice keyvalue store paper
Date Tue, 25 Oct 2011 22:13:36 GMT
No.  The problem is that you want to emulate real-world behavior which is
probably closer to 1000 threads each doing a single transaction and yielding
than anything else.  For instance, if your traffic originates from a
web-farm, each transaction will be in a thread that yields when it finishes
the current request.  That probably will only result in a few transactions
and each transaction will be handled by a region server in a different
thread.  How this reflects to the threads that actual touch the map being
considered is an interesting question, but there are likely to be a hundred
or more such threads.

Inserts, on the other hand, are much more likely to come from a single
thread doing lots of inserts but each insert will be a separate request and
thus be handled by different threads on the hbase regionserver side.

The problem that this causes is that you can blow cache as the access to the
table is handed from core to core and socket to socket.

On Mon, Oct 24, 2011 at 11:14 PM, Akash Ashok <thehellmaker@gmail.com>wrote:

> Wouldn't it be enough if I increase the number of iterations by a factor if
> say 100 per thread ?
>

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