Brian Akins wrote:
Bill Stoddard wrote:
Like I said, it can fill up two Gig pipes on a dual Opteron with ~25% idle time. mod_cache can't fill up one.
I'd certainly be interested in knowing how much faster mod_url_cache is as compared to mod_cache/mod_mem_cache. The only way I can see to -dramatically- improve the performance of mod_cache/mod_mem_cache is to completely bypass i/o filters. I have some patches to mod_mem_cache laying around that give perhaps a 10% performance improvement by prebuilding headers and aligning them on cache boundaries, but 10% is not really that great.
If you run in quick_handler, the only way you get filters (other than core_ouput) is if you add them your self. Guess if I did or not :)
If the headers are stored and "thawed" in an efficient manner (both processor and memory) they don't affect the performance much. Also, sometimes mmap is slower than just plain old read.
Also, using a filesystem and sendfile is much faster than serving from RAM, plus multiple processes can "share" the same cache. mod_cache doesn't do that.
All of these are based on my observartions of our environment.
Thanks, that's what I wanted to know. I can pretty much guess the implementation (I've done similar stuff before). mod_mem_cache takes a middle of the road approach; performance is sacrificed to provide some ability to filter i/o. However, it also makes sacrifices in functionality (like access/auth checking) by using the quick_handler to serve pages. Too many compromises for the 'i want full functionality' folks and too many compromises for the 'i want best raw performance' folks. mod_mem_cache tries to be all things to all masters but ends up falling short by all accounts. I've got two more patches lying around that I may dust off if i get any free time :-) The first introduces a config directive that enables moving the mod_mem_cache handler to the traditional handler phase (to pick up full access/auth checking and full filter capabilities). The second patch (I call it my 'go fast patch') does some experiments with prebuilding/caching the response headers and bypassing output filtering (in concept not unlike your mod_url_cache I expect).
Bill
