Messages in this thread | | | Date | Fri, 21 May 1999 14:23:26 +0100 (GMT) | From | Matthew Kirkwood <> | Subject | Re: andrea buffer code (2.2.9-C.gz) |
| |
On Fri, 21 May 1999, Andrea Arcangeli wrote:
> >Indeed. In cases like these, might it make some sense to have some fixed > >number of spinlocks and choose which to use with some reductive function? > > Impacting runtime performances to SMP scale better, is going to lose > according to me. I think that if you want to SMP scale better it worth > to pay _only_ with memory wastage and _not_ with wasted CPU cycles.
If you can be fairly sure that there won't be seriously degenerate situations, a simple reduction (a pair of xors to reduce 32bits to 256) wastes only 1K, and surely might improve performance on 4+-way machines?
> But as just said masp0008 raised the issue of the io_request_lock that > is just held all the time so we can't scale better at all even using one > spinlock per bit of memory in the machine :-). This is the only thing > that obviously convinced me to remove the spinlock immediatly.
Yes, that makes rather a big difference :
Matthew.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |