From: Miklos S. <mi...@sz...> - 2012-04-19 10:28:14
|
Pradeep Shetty <she...@gm...> writes: > Hello, > > I was working on fuse stuff and noticed that its random read performance is > almost 10 times > slower than the in-kernel FS. I ran a simple filebench random read > benchmark with and with > out pass-through fuse on ext4 running on Flash SSD device. It was a single > process, 10 threads - benchmark > and the performance difference is 10 times. In-cache performance is almost > the same as > the ext4. I would expect a slow down of 2-3 times but 10 times is little > too much for > the out of cache workload. One other thing is that when I make the > benchmark single threaded, > the fuse performance is better - I could see more IOs at the disk and it is > around 5-6 times slower > now instead of 10 times. It looks like it is getting serialized and also > loosing some performance > because of some lock contention and was wondering if any of you had come > across this and > know the issue. First guess is that it's caused by the readahead code messing up detection of random read pattern. Do you see similar slowdowns for sequential reads? Can you post instructions to reproduce? Thanks, Miklos |