From Bugzilla Helper: User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.3) Gecko/20030312 Description of problem: NFS write performance on ia64 is 5-7 times worse than ia32. Test boxes: ia32 - 1Ghz P3 x330, ia64 - 900Mhz RX6000, ia64 - 1.3Ghz RX6000 All tests are done using the same nfs fileserver. Tests were done sequentially. Fileserver and clients had no load at the time. This is easily reproduceable cd /nfs/<somewhere> 2.4.9 � ia32 ia32> time dd if=/dev/zero of=ddtest bs=1024 count=500000 500000+0 records in 500000+0 records out 0.270u 4.190s 0:53.82 8.2% 0+0k 0+0io 115pf+0w 2.4.20 � ia32 ia32> time dd if=/dev/zero of=ddtest bs=1024 count=500000 500000+0 records in 500000+0 records out 0.300u 4.610s 0:48.13 10.2% 0+0k 0+0io 115pf+0w 2.4.18 � ia64 - McKinely ia64> time dd if=/dev/zero of=ddtest bs=1024 count=500000 500000+0 records in 500000+0 records out 0.940u 9.125s 7:00.36 2.3% 0+0k 0+0io 73pf+0w 2.4.18 � ia64 - Madison ia64> time dd if=/dev/zero of=ddtest bs=1024 count=500000 500000+0 records in 500000+0 records out 0.247u 8.161s 5:30.27 2.5% 0+0k 0+0io 73pf+0w Version-Release number of selected component (if applicable): 2.4.18-e.37smp How reproducible: Always Steps to Reproduce: 1. In the description field 2. 3. Actual Results: ia64 shows at least 5x performace loss compared to ia32. Expected Results: Comparable results between the platforms would be expected. Additional info:
It seems to be related to the packet size. >= 16k packets yield comparable performance to ia32. Should the kernel ( or autofs ) default to something more reasonable on ia64?
kernel has no default; mount sets a default fwiw.
The problem is you are doing sync writes instead of async writes since the page size (16k)(on an ia64) is larger that server's write size which is typically 8k.