Given that goal, testing memcached against spcached is even less likely to tell you much, because differences between the way the two applications work internally may well be much more significant than any performance differences introduced by the particulars of the threading model. So if you see one of them is faster than the other, all you can really conclude is that one of them is faster than the other, but you can't infer anything about whether a half-sync/half-async thread pool is good or bad for performance. Various people have spent quite a lot of effort optimizing memcached, and I suspect spcached hasn't had similar attention paid to its internal bottlenecks and performance characteristics.
Just my opinion. (And sorry I didn't realize your goal or I would have replied to this over on the libevent list instead.) Well, actually, to make this relevant to the memcached list while further illustrating what I mean, I will tell people a few of the things we (Facebook) tried and ended up rejecting. We tried replacing the hashtable with a Judy tree and with a trie; both of those ended up being slower than the hashtable by a fair amount (theoretically a trie should be faster given that it's not doing the math to compute a hash value, but it turned out to be slower because walking the trie involved more CPU cache misses.) My initial threading implementation was a worker-pool model that farmed out individual requests to worker threads, but the poor interaction between threads and libevent caused us to make three or four extra system calls per request which again killed performance and drove CPU consumption way up. The overall point being that memcached -- while there is undoubtedly still plenty of room for improvement -- is tuned to the point where adding a few system calls per request, or slightly increasing the number of out-of-cache memory accesses, has a measurable impact on throughput. Unless you're measuring it against another server that has been similarly tuned, any difference you see is more likely to be due to that tuning than to any inherent difference between the two programs. -Steve liusifan wrote: > Hi, Steven Grimm > > I agree with you. I sorry for not to explain my purpose very clear. > > My purpose is to compare single-threaded event-driven program > vs. half-sync/half-async thread pool program. The both are based on libevent. > > Best regards， > > liusifan > 2007-05-18 > >
