On Fri, Mar 10, 2006 at 09:43:04PM +0800, Fuxin Zhang wrote:
> Thiemo Seufer 写道:
> > Can you provide information about the performance impact? That is,
> > what system(s) you use, and some basic benchmark numbers. Did you also
> > try 64k pages?
> The system is running debian sarge,with a mips like Chinese 4-way
> superscalar CPU and marvell 64240 system bridge, 256M sdram.Maybe
> related information: this CPU has 64K 4-way dcache/icache，so cache
> alias exists with 4K configuration.
>
> Big page size helps a bit with at least computational programs, such as
> SPEC CPU2000 benchmarks. The attached jpg is a graph showing the results
> of some SPEC programs under different page sizes. But we have not
> benchmarked other workloads. 16K page size is used mainly to avoid cache
> alias, which caused us too much time.
>
> Yes, with 64k more patches needed(some "short" field overflow?I don't
> remember the exact modifications,other guys did it)
There is more to large pages than just improved performance. They'll
also save 75% memory for mem_map, will eleminate cache aliases in a
bulletproof way. And will make allocation of memory larger than 4kB
reliable - this especially is a big bonus with ethernet jumboframes with
some NICs.
Ralf
|