From: "Chris Wedgwood" <cw@f00f.org>
> I tried various things -- none changed the results significantly. In
> the end I had both reader and writer doing buf[0]++ figuring and lazy
> mappings or something would have to be trigger buy that.
Could you boot with profiling enabled (profile=1 on command line) and check
readprofile?
The Linux pipe code has 2 problems:
* it uses wake-all, and a semaphore: if you have multiple readers and
writers, then you have an incredible amount of idle reschedules. This is no
problem for pipes, but could be bad for fifo's. I wrote a patch, but Linus
decided that this is to rare to optimize for.
* it needs one task switch for 4 kB data transfer.
On Mon, Mar 13, 2000 at 02:04:38AM -0800, David S. Miller wrote:
>
> FreeBSD's pipe code plays vm tricks to not even copy the
> data from one process to another.
>
Do you know how long a typical pipe write is? FreeBSD plays vm tricks for
writes > 8 kB.
I know that page table changes would kill SMP performance, but it might be
possible to use a single-copy pipe transfer with kiovec's without any page
table changes. [current code: 2 copies + lots of schedules]
./bw_pipe writes 64kB blocks, and thus linux needs 17 syscalls and 32
context switches for each 64k block.
one "./bw_pipe" gives me 75000 context switches...
--
Manfred
-
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/
This archive was generated by hypermail 2b29 : Wed Mar 15 2000 - 21:00:30 EST