On Wed, Sep 7, 2011 at 2:09 PM, Hefeng Yuan <hfyuan@rhapsody.com> wrote:
We didn't change MemtableThroughputInMB/min/maxCompactionThreshold, they're 499/4/32.
As for why we're flushing at ~9m, I guess it has to do with this: http://thelastpickle.com/2011/05/04/How-are-Memtables-measured/
The only parameter I tried to play with is the compaction_throughput_mb_per_sec, tried cutting it in half and doubled, seems none of them helps avoiding the simultaneous compactions on nodes.

I agree that we don't necessarily need to add node, as long as we have a way to avoid simultaneous compaction on 4+ nodes.


Can you check in the logs for something like this 
...... Memtable.java (line 157) Writing Memtable-<ColumnFamilyName>@1151031968(67138588 bytes, 47430 operations)
to see the bytes/operations at which the column family gets flushed. In case you are hitting the operations threshold you can try increasing that to a high number. The operations threshold is getting hit at  less than 2% of size threshold. I would try bumping up the memtable_operations substantially. Default is 1.1624999999999999(in millions).  Try 10 or 20 and see if your CF flushes at higher size. Keep adjusting it until the frequency/size of flushing becomes satisfactory and hopefully reduces the compaction overhead.


On Sep 7, 2011, at 10:51 AM, Adi wrote:

On Wed, Sep 7, 2011 at 1:09 PM, Hefeng Yuan <hfyuan@rhapsody.com> wrote:

The reason we're attempting to add more nodes is trying to solve the long/simultaneous compactions, i.e. the performance issue, not the storage issue yet.
We have RF 5 and CL QUORUM for read and write, we have currently 6 nodes, and when 4 nodes doing compaction at the same period, we're screwed, especially on read, since it'll cover one of the compaction node anyways. 
My assumption is that if we add more nodes, each node will have less load, and therefore need less compaction, and probably will compact faster, eternally avoid 4+ nodes doing compaction simultaneously.

Any suggestion on how to calculate how many more nodes to add? Or, generally how to plan for number of nodes required, from a performance perspective?


Adding nodes to delay and reduce compaction is an interesting performance use case :-)  I am thinking you can find a smarter/cheaper way to manage that.
Have you looked at 
a) increasing memtable througput
What is the nature of your writes?  Is it mostly inserts or also has lot of quick updates of recently inserted data. Increasing memtable_throughput can delay and maybe reduce the compaction cost if you have lots of updates to same data.You will have to provide for memory if you try this. 
When mentioned "with ~9m serialized bytes" is that the memtable throughput? That is quite a low threshold which will result in large number of SSTables needing to be compacted. I think the default is 256 MB and on the lower end values I have seen are 64 MB or maybe 32 MB.

b) tweaking min_compaction_threshold and max_compaction_threshold
- increasing min_compaction_threshold will delay compactions
- decreasing max_compaction_threshold will reduce number of sstables per compaction cycle
Are you using the defaults 4-32 or are trying some different values

c) splitting column families
Again splitting column families can also help because compactions occur serially one CF at a time and that spreads out your compaction cost over time and column families. It requires change in app logic though.