latencies due to disk writes

latencies due to disk writes

[kalium]

Hello!

I'm experiencing extreme lags during disk writes. I have read somewhere (didn't save the URI, sigh) that this is actually related to bad (non-existing) write io priorities (CFQ only manages file reads).

I could imagine two quick, easy and probably quite effective ways to prevent such lags:

1.) don't flush buffers to disk at once more than necessary.

2.) relate CPU niceness to max write buffer fill level (ie. the point where it gets forced to be flushed to disk -- a conservative estimate would be much better than nothing): (100-5*nicelevel)%, ie. writes for processes having nice level 19 are blocked/delayed until the write buffer is below 5%. That way, the accounting is done at a higher and probably easier to access level.

Maybe I'm just talking nonsense, but nonetheless, here are my 2 cents.

Best regards,
Mark

p.s. please CC me as I'm not subscribed to this list.
-- 
"Feel free" - 10 GB Mailbox, 100 FreeSMS/Monat ...
Jetzt GMX TopMail testen: www.gmx.net/de/go/mailfooter/topmail-out

Re: latencies due to disk writes

[Frank Ch. Eigler]

kalium@gmx.de writes:

> I'm experiencing extreme lags during disk writes. [...]

Wacky idea: "ionice -n 7 -c 2" down the pdflush/swapd kernel threads.

- FChE

Re: latencies due to disk writes

[Mark Stier]

> > I'm experiencing extreme lags during disk writes. [...]
>
> Wacky idea: "ionice -n 7 -c 2" down the pdflush/swapd kernel threads.

As I already wrote in my post, no disk IO scheduler cares for disk
writes. So even CFQ is only a disk *IN* scheduler...

Re: latencies due to disk writes

[Bill Davidsen]

kalium@gmx.de wrote:
> Hello!
> 
> I'm experiencing extreme lags during disk writes. I have read somewhere (didn't save the URI, sigh) that this is actually related to bad (non-existing) write io priorities (CFQ only manages file reads).
> 
> I could imagine two quick, easy and probably quite effective ways to prevent such lags:
> 
> 1.) don't flush buffers to disk at once more than necessary.
> 
Actually, in many cases this is just what you do want, to avoid filling 
memory with buffered writes and then flushing them on time or memory runout.

Investigate the /proc/sys/vm/dirty_* values.

> 2.) relate CPU niceness to max write buffer fill level (ie. the point where it gets forced to be flushed to disk -- a conservative estimate would be much better than nothing): (100-5*nicelevel)%, ie. writes for processes having nice level 19 are blocked/delayed until the write buffer is below 5%. That way, the accounting is done at a higher and probably easier to access level.
> 
> Maybe I'm just talking nonsense, but nonetheless, here are my 2 cents.
> 
> Best regards,
> Mark
> 
> p.s. please CC me as I'm not subscribed to this list.

-- 
Bill Davidsen <davidsen@tmr.com>
   "We have more to fear from the bungling of the incompetent than from
the machinations of the wicked."  - from Slashdot