If its the RMI horror ( registry port + actual port ) you are worried about, enabling MX4J in cassandra-env.sh for the HTTP transport ( http://wiki.apache.org/cassandra/Operations#Monitoring_with_MX4J ) could help although i never used it myself. Or you can just throw in the Jolokia JVM agent http://www.jolokia.org which will do roughly the same

i'm not aware of the hit rate or any of the other cf/system metrics being exposed over anything other than JMX, sorry. You can poke around with nodetool on the nodes themselves ( that uses JMX itself ) to get other metrics, but i have not been able to get cache stats out of nodetool cfstats. I think it might be because they are no longer on the CF's mbean.

regards,
Andras

On 27 Nov 2012, at 05:22, Yiming Sun <yiming.sun@gmail.com> wrote:

Interesting -- I completely overlooked the OS cache aspect, but how long does the OS keep cached pages? It seems they reside in memory for 4 hours at least (which happens to be the default time for key cache)..  The CF has over 2.6 million rows, and I am only fetching a random set of 5000 each time. 

Where else can I examine the cache hit ratio? The nodes are remote servers behind some strict university controlled firewall, so it wouldn't be easy for me to tap in using JConsole.  Thanks.

-- Y.


On Mon, Nov 26, 2012 at 11:06 PM, Andras Szerdahelyi <andras.szerdahelyi@ignitionone.com> wrote:

SSTables in http://en.wikipedia.org/wiki/Page_cache maybe?
How many rows do you have in this CF? Are you getting all columns?

What do the cassandra.db mbeans say ( hit ratio, cache requests, items etc )

regards,
Andras

On 27 Nov 2012, at 04:12, Yiming Sun <yiming.sun@gmail.com> wrote:

> Hi,
>
> I am carrying out some performance test against a 6-node cassandra cluster (v1.1.0), and need to disable the key cache entirely as one of the scenarios.
>
> However, by setting key_cache_size_in_mb to 0, I am still seeing caching effects.  For example, when I fetch a set of 5000 rows, the first time it would deliver at about 60MB/sec, but all subsequent fetches (even after fetching several other sets), it would deliver at about 120MB/sec.  This is exactly the same behavior I have observed when the key_cache_size_in_mb was left blank.
>
> So I am just wondering how I can disable the key cache entirely.  All our nodes have JNA.jar deployed -- not sure it this has anything to do with what I saw.  Thanks.
>
> -- Y