9fans archive / 1999 / 02 / 7 /    prev next

From: Franklin Robert Araujo França 973930@dcc...
Subject: [9fans] Comments on File I/O measurements
Date: Wed, 03 Feb 1999 10:14:20 +0000

 Russ, please note below our corrections to your random read
calculations.

Comments on File I/O measurements
The first test (bw_file_rd) reads an 8MB file into memory
(the file should then be cached since main memory is 64MB),
then measures the time for re-reading the file sequentially.
The other tests read and write a 100 MB file in 8KB blocks,
using the lmdd test from lmbench.
We were amazed by the Linux numbers so we fetched the Quantum
ST32AT specifications from www.quantum.com; it turns at
5400 rpm, average seek time=10ms,has a 128 KB internal buffer
(~1 track) and a variable number of 512 byte sectors/track
( from 277 to 154) giving nominal rates varying from
12.2 to 6.8 MB/s. Linux partition is after the first 25% of
the disk size, so we estimated nominal rate at that point at
9.2 MB/s. We rebooted the system before all measurements in order
to avoid any cache re-use, and executed the sync() system call
before ending the time measurements.
Linux was able to create and write sequentially a 100MB file
at 50% of the disk nominal speed and (after reboot) to read
sequentially the same file at 90% of the disk nominal speed.
In the random read test with the same file we read 12800 times
8KB blocks randonly chosen (therefore, 12800 random lseeks
followed by read): the time measured was 116.3 secs giving
9.8 ms /operation, very close to the average seek time.
PS: Plan9 partition is after Linux in the disk, which may
give nominal rates up to 20% lower than Linux 9MB/s, but this
obviously cannot explain the difference in performance.