Re: [PATCH] cacheline align pagevec structure

[Marcelo Tosatti]

On Thu, Sep 09, 2004 at 03:49:06PM -0700, Andrew Morton wrote:
> Marcelo Tosatti <marcelo.tosatti@xxxxxxxxxxxx> wrote:
> >
> > Right now it is 140 bytes on 64-bit and 72 bytes on 32-bit. Thats just a little bit more 
> > than a power of 2 (which will cacheline align), so shrink it to be aligned: 64 bytes on 
> > 32bit and 124bytes on 64-bit. 
> > 
> > It now occupies two cachelines most of the time instead of three. 
> > 
> > I changed nr and cold to "unsigned short" because they'll never reach 2 ^ 16.
> > 
> > I do not see a problem with changing pagevec to "15" page pointers either, 
> > Andrew, is there a special reason for that "16"? Is intentional to align
> > to 64 kbytes (IO device alignment)? I dont think that matters much because
> > of the elevator which sorts and merges requests anyway?
> > 
> > 
> > 
> > Did some reaim benchmarking on 4way PIII (32byte cacheline), with 512MB RAM:
> > 
> > #### stock 2.6.9-rc1-mm4 ####
> > 
> > Peak load Test: Maximum Jobs per Minute 4144.44 (average of 3 runs)
> > Quick Convergence Test: Maximum Jobs per Minute 4007.86 (average of 3 runs)
> > 
> > Peak load Test: Maximum Jobs per Minute 4207.48 (average of 3 runs)
> > Quick Convergence Test: Maximum Jobs per Minute 3999.28 (average of 3 runs)
> > 
> > #### shrink-pagevec #####
> > 
> > Peak load Test: Maximum Jobs per Minute 4717.88 (average of 3 runs)
> > Quick Convergence Test: Maximum Jobs per Minute 4360.59 (average of 3 runs)
> > 
> > Peak load Test: Maximum Jobs per Minute 4493.18 (average of 3 runs)
> > Quick Convergence Test: Maximum Jobs per Minute 4327.77 (average of 3 runs)
> 
> I think the patch make sense, but I'm very sceptical about the benchmarks
> ;)

Why's that? You think changing to the number of pages in the pagevec to "15" instead
"16" is the cause?

Or that the performance increase is not a direct effect of the one cacheline 
saved per pagevec instance?

Or you think such benchmark is too specific to be interpreted as a broad
vision of performance? 

:)

-
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/