Re: [gentoo-user] Re: Best way to improve interactivity with heavy disk activity?

[Mark Knecht]

On Mon, Dec 6, 2010 at 7:58 AM, Nikos Chantziaras <realnc@×××××.de> wrote:
> On 12/06/2010 04:33 PM, Stroller wrote:
>>
>> On 27/11/2010, at 10:22pm, Nikos Chantziaras wrote:
>>>>
>>>> ...
>>>> Does using the ck-sources kernel reduce the necessity to use ionice?
>>>
>>> To some degree, yes.
>>
>> Tried this late last week. You chose your words perfectly - there is only
>> some degree of improvement.
>>
>> The machine seems much snappier if I return to an interactive session
>> after leaving it idle for some time, but during DVD rips it is still
>> unresponsive for tens of seconds at a time. It seems like maybe it responds
>> quicker than mainline sources, but it's still so slow that it's both
>> unusable and hard to be sure whether that's the case.
>>
>> Will try cgroups this week, perhaps.
>
> cgroups will not help with that either.  I bet the problem is that the
> kernel doesn't care about what kind of data it is caching.  I came across
> this on LKML, but no one was interested in a solution.
>
> The problem is that the kernel caches all data.  Ripping a DVD will consume
> 4GB cache, and all other data will be thrown out.  Throwing that all out
> again after the rip has finished is an awful lot of I/O load. The smart
> thing to do would be for the kernel to not cache such stuff (DVD rips, big
> video files, etc.)  But it does.

I forgot to ask Stroller if he's looked at top while these delays
happen. If the issue is writing to disk from cached memory then what I
saw when that was occurring (and during the times my system was
unresponsive) was very low CPU usage while having very high wait
percentages.

In my 6 core/12 thread top display I'd see 2 or 3 CPUs involved in the
wait (writing data out but delayed by responses from the disk driver I
suppose) while the other CPUs were essentially idle, presumably
because they couldn't get to disk for whatever they needed.

- Mark