it looks to me like you are describing "compaction causes a lot of i/o." see http://wiki.apache.org/cassandra/MemtableSSTable#Compaction
things you can do about the extra i/o: - increase memtable sizes (if you haven't done this yet you should, by 10x or so) - reduce compaction priority (see http://www.riptano.com/blog/cassandra-annotated-changelog-063) - enable the dynamic snitch (see http://www.riptano.com/blog/whats-new-cassandra-065) so other nodes can route around one that is slow during a compaction you can't make it stop using extra space during the compaction itself, that is part of the design (and part of the price you pay for not doing random i/o at insert time). On Wed, Sep 8, 2010 at 2:08 PM, Dathan Pattishall <datha...@gmail.com> wrote: > This link describe ganglia / cassandra graphing. > > http://mysqldba.blogspot.com/2010/09/cassandra-and-ganglia.html > > I ran into a problem illustrated here. > > http://www.flickr.com/photos/dathan/4971255111/ > > This screen shot shows a huge spike of transport exceptions between the > hours of 12:15 - to 1:30. Why? Lets see. > > http://www.flickr.com/photos/dathan/4971869002/ > > This link shows that the pending reads jump because the message > deserialization pool (mutex) blocks or maybe its viceversa. But Why? Lets > see. > > This link shows that wait_io on the box sky rocketed. > > http://www.flickr.com/photos/dathan/4971290101/ > > but why? > > Could it be because > > http://www.flickr.com/photos/dathan/4971869054/ > > This graph shows a massive amount of data growth for this server, then it > reduces but why? How can I tune it so that a growth of data doesn't explode > like this? > > > > Some background information: > > These servers are DELL 2950 dual quad core boxes with 48GB of Ram on a > RAID-10 EXT3 FS backed by 8 disks on a PERC-6 Controller with BBC. Each > server rougly recieves 300-400 requests per second fronted by a F5 > Loadbalancer (soon to be HA-Proxy) on least connections, doing a client stat > check to verify the server is up from a client point of view. > > There is only one simple key space. A Super Column is defined but not used > and uses a RandomPartitioner with NO RowCaching and mmap enabled. > > > > > > > > > > > > > -- Jonathan Ellis Project Chair, Apache Cassandra co-founder of Riptano, the source for professional Cassandra support http://riptano.com
