Thread

Would ext3 data=journal help for Postgres synchronous io mode

Yusuf Goolamabbas <yusufg@outblaze.com> — 2002-09-27T02:55:10Z

According to ext3 hackers (Stephen Tweedie, Andrew Morton). ext3
data=journal mode is much faster than any of the other mode for
workloads which do a lot of syncrhonous i/o. Personally, I have seen
dramatic improvements on moving mail queues to this mode (postfix in
particularly flies with this mode)

While this may seem contradictory (forcing journaling for the data in
addition to the metadata), it will likely improve the performance for
sync I/O loads like mail servers because it can do all of the I/O to the
journal without any seek or sync overhead while the mail is arriving.

I assume that since Postgresql does a lot of fsyncs, it would benefit
also. I have sent email to Sridhar asking if he could test this

Another thing to note is that Linux 2.4.x kernels < 2.4.20-pre4 use
bounce buffer's to do IO if the machine has > 1GB memory. Distributor
kernels such as Redhat/Suse/Mandrake are patched to do IO via DMA
to/from highmem (>1GB). According to IBM's paper @ OLS, this improves IO
performance by 40%

BTW, Is this list archived on the website

Regards, Yusuf
-- 
Yusuf Goolamabbas
yusufg@outblaze.com