OK, Here is some profiling data from Samba. I profiled around each of the GetTimeOfDay calls in construct_reply. So, it does not count all gettimeofday calls. I was more interested in the fastpath.
I then ran a test that read a single 50MB file from the server. The file would pretty much be in cache as I wrote it there a few minutes before and I have 2GB on the server. Here are the non-zero syscall values: opendir_count: 1 opendir_time: 113 readdir_count: 13 readdir_time: 55 rmdir_count: 1 rmdir_time: 39 closedir_count: 1 closedir_time: 62 open_count: 2 open_time: 37 close_count: 2 close_time: 18 read_count: 833 read_time: 144950 stat_count: 8 stat_time: 160 fstat_count: 833 fstat_time: 3023 chdir_count: 1 chdir_time: 28 fcntl_lock_count: 833 fcntl_lock_time: 1493 gettimeofday_count: 852 gettimeofday_time: 1527 All times are in microseconds. Firstly, my earlier assertion that gettimeofday is about as costly as fstat is off by about a factor of two. Secondly, this code includes the sendfile patch, and the read times are all the time attributed to sendfile. The overall cost of gettimeofday seems to be less than 1% here and would be smaller in a more realistic mix. Regards ----- Richard Sharpe, [EMAIL PROTECTED], [EMAIL PROTECTED], [EMAIL PROTECTED]