Re: RFC: dirty_ratio back to 40%

From: Zan Lynx
Date: Thu May 20 2010 - 20:50:11 EST


On 5/20/10 5:48 PM, KOSAKI Motohiro wrote:
Hi

CC to Nick and Jan

We've seen multiple performance regressions linked to the lower(20%)
dirty_ratio. When performing enough IO to overwhelm the background
flush daemons the percent of dirty pagecache memory quickly climbs
to the new/lower dirty_ratio value of 20%. At that point all writing
processes are forced to stop and write dirty pagecache pages back to disk.
This causes performance regressions in several benchmarks as well as causing
a noticeable overall sluggishness. We all know that the dirty_ratio is
an integrity vs performance trade-off but the file system journaling
will cover any devastating effects in the event of a system crash.

Increasing the dirty_ratio to 40% will regain the performance loss seen
in several benchmarks. Whats everyone think about this???

In past, Jan Kara also claim the exactly same thing.

Subject: [LSF/VM TOPIC] Dynamic sizing of dirty_limit
Date: Wed, 24 Feb 2010 15:34:42 +0100

> (*) We ended up increasing dirty_limit in SLES 11 to 40% as it used to be
> with old kernels because customers running e.g. LDAP (using BerkelyDB
> heavily) were complaining about performance problems.

So, I'd prefer to restore the default rather than both Redhat and SUSE apply exactly
same distro specific patch. because we can easily imazine other users will face the same
issue in the future.

On desktop systems the low dirty limits help maintain interactive feel. Users expect applications that are saving data to be slow. They do not like it when every application in the system randomly comes to a halt because of one program stuffing data up to the dirty limit.

The cause and effect for the system slowdown is clear when the dirty limit is low. "I saved data and now the system is slow until it is done." When the dirty page ratio is very high, the cause and effect is disconnected. "I was just web surfing and the system came to a halt."

I think we should expect server admins to do more tuning than desktop users, so the default limits should stay low in my opinion.

--
Zan Lynx
zlynx@xxxxxxx

"Knowledge is Power. Power Corrupts. Study Hard. Be Evil."
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/