2001-10-18 20:54:19

by Zlatko Calusic

[permalink] [raw]
Subject: Write throughput in >= 2.4.10

It looks like recent kernels have some serious trouble during simple
writing of files. Throughput is cut to half.

2.4.12-ac3 (Riel VM):

procs memory swap io system cpu
r b w swpd free buff cache si so bi bo in cs us sy id
0 4 1 0 2548 7652 369104 0 0 0 21992 291 592 0 23 77
0 4 1 0 2556 7652 369104 0 0 0 22500 280 175 0 3 97
0 4 1 0 3064 7652 368588 0 0 8 19644 278 202 0 4 96
...

2.4.13-pre4 (Andrea/Linus VM):

procs memory swap io system cpu
r b w swpd free buff cache si so bi bo in cs us sy id
1 1 1 30312 2884 2304 384076 0 28 0 7784 199 241 0 12 88
0 1 1 30316 2068 2256 385772 0 132 4 7900 186 188 1 8 91
0 1 0 30316 3960 2232 384140 0 0 8 6744 179 204 0 4 96
...

'bo' column is the one to check out... I copied just 3 lines, but they
are all alike. 2.4.13-pre ends up with 11MB/sec, where -ac kernels are
over 20MB/sec (during sequential writing of big files - ext2 of course).

Also it looks like pre4 swaps when it is not necessary to do so. With
380MB in page cache I don't expect any swap traffic at all.
--
Zlatko


2001-10-18 20:57:59

by Marcelo Tosatti

[permalink] [raw]
Subject: Re: Write throughput in >= 2.4.10



On 18 Oct 2001, Zlatko Calusic wrote:

> It looks like recent kernels have some serious trouble during simple
> writing of files. Throughput is cut to half.
>
> 2.4.12-ac3 (Riel VM):
>
> procs memory swap io system cpu
> r b w swpd free buff cache si so bi bo in cs us sy id
> 0 4 1 0 2548 7652 369104 0 0 0 21992 291 592 0 23 77
> 0 4 1 0 2556 7652 369104 0 0 0 22500 280 175 0 3 97
> 0 4 1 0 3064 7652 368588 0 0 8 19644 278 202 0 4 96
> ...
>
> 2.4.13-pre4 (Andrea/Linus VM):
>
> procs memory swap io system cpu
> r b w swpd free buff cache si so bi bo in cs us sy id
> 1 1 1 30312 2884 2304 384076 0 28 0 7784 199 241 0 12 88
> 0 1 1 30316 2068 2256 385772 0 132 4 7900 186 188 1 8 91
> 0 1 0 30316 3960 2232 384140 0 0 8 6744 179 204 0 4 96
> ...
>
> 'bo' column is the one to check out... I copied just 3 lines, but they
> are all alike. 2.4.13-pre ends up with 11MB/sec, where -ac kernels are
> over 20MB/sec (during sequential writing of big files - ext2 of course).
>
> Also it looks like pre4 swaps when it is not necessary to do so. With
> 380MB in page cache I don't expect any swap traffic at all.

Zlatko,

Could you please try 2.4.12-pre3 instead pre4 ?

Linus has made some changes and I want to see if those are partly
responsible for the problems you're seeing.

Thanks


2001-10-18 21:26:12

by Zlatko Calusic

[permalink] [raw]
Subject: Re: Write throughput in >= 2.4.10

Marcelo Tosatti <[email protected]> writes:

> On 18 Oct 2001, Zlatko Calusic wrote:
>
> > It looks like recent kernels have some serious trouble during simple
> > writing of files. Throughput is cut to half.
> >
[snip]

>
> Zlatko,
>
> Could you please try 2.4.12-pre3 instead pre4 ?
>
> Linus has made some changes and I want to see if those are partly
> responsible for the problems you're seeing.
>

pre3 behaves exactly the same, so pre4 is not the culprit.
--
Zlatko