Hello, I'm running two GlusterFS nodes for a replicated data storage setup trough a high latency DSL line (60-70ms RTT). File access works alright, i.e. files are obviously accessed on the local node, but all operations involving meta data (such as directory listings) are incredibly slow (20 seconds and more for a well-populated directory of ~200 entries). The link is fairly low bandwith, too (128Kbit/s), but I simulated that situation on the high bandwith link of my test setup using traffic shaping and still got a rather snappy response. Testing happening late yesterday night I wasn't really up to figuring out how to worsen latency using queue disciplines. But seeing as latency is the only difference between the test setup and the production environment (0.3-0.5ms vs. 60-70ms) it seems to be the cause of the problem. I tried read-ahead with a cache and and the read-subvolume option, but to no avail. Both only seem to apply to the data payload of files, i.e. actually reading data from a file works at the speed you'd expect from local storage but metadata access remains sluggish. Are there any more knobs one could fiddle with to accomodate the high latency line? If not would a reduction of the latency to 10-20ms RTT help (the DSL lines involved could probably be switched to Fastpath operation)? Cheers, Johannes -- In the begining was the word, and the word was: Content-Type: text/plain; charset=us-ascii
