Dear Henrik Nordstrom: I had get the final result of comparing three version.In my new test environment, the response time: orig 14.227ms no_pipe 13.550ms pipe_patch 13.905ms I have one questions below.
======= 2004-02-17 12:26:00 you wroted :======= >On Tue, 17 Feb 2004, [GB2312] 夏洪涛 wrote: > >> Oh, I get it. Do you mean: When poll/select syscall wait on the ready >> FDs, the pipe can wake up poll/select syscall, and enter another loop. >> At each loop, storeDirCallback() will be called first. It will poll the >> disk I/O status. > >Yes. > >> 1. Dose poll/select never check FDs that represent opened files? They >> can not poll the the disk I/O status themselves? > >poll/select can not check the readiness of local files. Such >filedescriptors will always indicate ready for reading/writing even if >such operations would block the process waiting for the disk. I read the code, and found that in disk.c, function diskHandleWrite() and diskHandleRead() had call commSetSelect() to poll the file descriptors. Then if poll/select can not check the readiness of local files, the purpose to call commSetSelect() is just divide write or read operation into pieces , so that other request can be processed. This can increase the throughput.Am I right? > >> 2. With high request ratio, the poll/select syscall seldom wait for >> FDs,the pipe mechanism seems do bad to performance.Is it right? > >The pipe mechanism can potentially degrade performance somewhat in high >request ratio as the context switching increases and more CPU is wasted, >but I did not find any noticeable such drawbacks when testing this when >the completetion pipe was implemented. > >The change I proposed in my previous email should address this by >allowing fallback on batch processing on high loads at the expense of >a risk of somewhat higher I/O latency. > >> I had test it roughly. The difference between two results seems not so >> obvious.I'll test it more. > >Please tell more. > >Regards >Henrik = = = = = = = = = = = = = = = = = = = = Best Regards Xia Hongtao [EMAIL PROTECTED] 2004-02-18