nate at verse.com
Wed Jan 15 21:22:09 PST 2014
Several months ago, Jason Evans wrote:
> Also of relevance: the SSD backing feature you added existed
> in 2.x versions of jemalloc, but I removed it because no one
> ever claimed to have found it useful.
I'm just learning about this now, but if it's what I want it to be,
this seems like a very useful feature. Do you have a description of
it written up, or a link to the code that last used it?
What I've been wanting for quite a while is a modern malloc() that
does all of it's allocations from one or a small number of arenas of
mmap() files. All the meta-information (free-lists, locking, etc.)
should be stored in the arena itself.
It would be used as a shared memory segment between processes. The
main use case would be holding index records for real time search.
There would be a limited number of writers but many readers.
The entry point for the reader processes would be at some known offset
that holds an index (or index of indexes). All the rest of
allocations would be would be accessible from there. New readers
could jump in at any point, old ones could die, but as long as the OS
doesn't crash at the wrong point the data would be synced to disk.
More information about the jemalloc-discuss