> In the context of the current state of the mem cache, yes, I do
> consider that a bad idea. The mem cache is not well tested. And in
> addition, on systems with good buffer caches, the disk cache will be
> able to share one memory/buffer cached object across all threads and
> processes and use zero-copy IO (sendfile) to deliver it to the
> network.
OK. So I will try more carefully the disk module.
My production platform will be Linux with 2.6 kernel.
I hope that Linux kernel buffer caches are able to manage properly objects with such big size.
>> I'm preparing a patch to correct the mem module bug, and writing a new
>> cache module based on shared memory.
>> Do you think that such new module will be useful for the community ?
> Possibly. You'd need to show that it actually outperformed the disk
> cache in some scenarios. It's a topic you should definitely bring up
> on the dev list rather than here.
I plan now to perform some benchmarks before to continue my development.
Thanks a lot for your help, Joshua.
Best regards,
Fred
|