Gershon,

Gershon Geva wrote:


I have a video recording system that writes 400 buffers per second, each ranging from 100 bytes to 4KB.


I am trying to tweak bdflush in order to optimize it for this specific usage profile. Can anyone suggest

What is the best strategy in terms of trade-off between flushing frequency and memory usage?

My main constraints are write speed – it must not exceed 1 second, and I want to keep the duty cycles

        Below 50% - meaning I must let the disk rest for 50% of the time.

BTW: The disk buffer size is 8MB, and theRAM size is 512MB.



You did not tell us if these writes are sequential or random. From the nature of your system (video recording) I assume it is sequential.

If so, I would try:

"50 500 0 0 60 100 80 0"

But bare in mind that this will allow the kernel less oppurtunties for joining writes. If you have several streams bweing recorded at once (and you probably have), this can be problematic. The only really good suggestion is to play with it a little and measure, measure, measure...

Other tricks to try:

Use the noatime option when mounting the file system.

If using ext3 try using data=journal (yes, I know it should be the slowest by common sense. Common sense is sometime over rated ;-)

Try SGI's xfs file system. it was designed with video in mind.

Re-design your application to do it's own flash by using O_DIRCT and AIO. Ok, this last one is a bit of a killer... ;-)

Some links that will help you:

http://www-106.ibm.com/developerworks/linux/library/l-fs8.html


Hope this helps, Gilad


-- Gilad Ben-Yossef <[EMAIL PROTECTED]> Codefidence. A name you can trust(tm) Web: http://codefidence.com | SIP: [EMAIL PROTECTED] Tel: +972.9.8650475 ext. 201 | Fax: +972.9.8850643

"I am Jack's Overwritten Stack Pointer"
        -- Hackers Club, the movie

=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]



Reply via email to