cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From karim duran <>
Subject Re: mmaped files and swap
Date Wed, 13 Mar 2013 16:52:55 GMT
I agree with Edward Capriolo,
Even when swap is enabled on your system, swaping rarely occurs on OS
today...(except for very loaded systems).

But, take care that some 32 bits system kernels allows only 2^32 bits
memory mapped file length ( ~ 2 Go ).
It could be a limitation for NoSQL databases. It's the case for MongoDB on
32 bits OS <>.

I don't know how to avoid swaping if Cassandra exceeds these limitation
when this case occurs.

2013/3/13 Edward Capriolo <>

> You really can not control what the OS-swaps out. java has other memory
> usage outside the heap, and native memory. best to turn swap off. Swap is
> kinda old school anyway at this point. It made sense when machines had 32MB
> RAM.
> Keeping your read 95th percentile low is mostly about removing deviations
> that cause requests to slow down, swap is one of the things that cause
> fluctuation becuase it is not predictable.
> On Wed, Mar 13, 2013 at 10:39 AM, Fredrik <
>> wrote:
>> I've got a question regarding understanding the recomendation to disable
>> swap.
>> Since Cassandra uses mlockall to lock the heap in RAM what is the reason
>> for disabling swap?
>> My guess is that is has to do with memory mapped files but as of my
>> understanding, accessing pages of
>> memory mapped files, those pages are never put in swap since they're
>> backed by files on disk and the OS
>> writes those pages to the memory mapped file instead of swap.
>> We've seen on Cassandra installations on Linux with swap enabled that
>> parts of the java process is swaped out and increasing.
>> So what's swaped out?
>> Regards
>> /Fredrik

View raw message