Re: Linux slowdown on 20M - summary

Marcin Dalecki (
Sat, 26 Apr 1997 11:14:25 +0200 (MET DST)

On Fri, 25 Apr 1997, Pavel Machek wrote:

> Hi!
> [I'm now connected via cslip on 2400 modem, so sorry for bad formating/
> typos]
> I recently posted results of some tests - consisting of compilation of
> 2.1.35 kernel on 2.1.35 linux. Everything on 486DX2/80 with
> 20MB of ram.
> On 20MB of ram, it took 32 minutes (128k l2 enabled)
> on 16MB of ram (or with mem=16M option) it took
> 22 minutes (with l2 cache enabled).
> Everyone told me that it is HW problem because my high 4MB of ram
> are not cacheable.
> So I turned level 2 cache off completely.
> COmpile time with mem=16M was 30 minutes,
> with 20MB of ram it was 35 minutes.
> So problem still exists with l2 disabled.
> [It still may be hw, however]
> Other think I noticed is that time from Uncompressing kernel to Elf:login differs heavily (15 sec X 25 sec) with 16MB and 20MB
> of memory (16MB is faster). Also 'interactive performance' is much
> worse with 20MB.
> Both even if l2 disabled.
> Other ideas (beside of cache problems were:)
> Problems with disk access not possible to memory higher that 16MB directly -
> impossible since I do not use DMA on disks and since disk activity when
> compiling kernel is almost none.
> David (?) Told me something about inode cache.
> Pavel

Yes. If You have more memmory the kernel doesnt try that hard to preserve
an good wirking set of pages in memmory. Therefore it does possible more
swapping on the execution of the compiler. But that shouldn't make an that
big difference. More noticable the pagetables hold by the kernel are
bigger in case of 20 MB of memory. (They are actually twice as big due to
the roundup from 20 to 16MB.) Therefore the corresponding tables (which
are used havly) May not fill that well anymore into the L1 chache and
adress lookup buffers of the processor. Causeing the whole bussystem to
slow down.