Re: optimized SHA1 (was Re: [Monotone-devel] net.venge.monotone.experiment.performance)

[Daniel Carosone]

On Sat, Aug 05, 2006 at 06:07:49PM +0200, Johan Bolmsj? wrote:
> One thing to keep in mind is that there are hardware accelerators for SHA1 
> out 
> there and I think OpenSSL has support for some of them. Wouldn't it be nice 
> if monotone had support for the hardware accelerators?

It might. There's an interesting tradeoff with hardware crypto
accelerators - they typically involve extra DMA round trips to memory
and syscall/interrupt processing overhead, instead of just chewing
through the work in the cpu and (hopefully) cache.

The result can sometimes be that the device adds latency and can take
*longer* for a straight-line operation.  Where they can add value is
on servers where you have other work for the cpu to do in the
meantime, and multiple parallel sessions so you can see a win on
aggregate throughput.  Even there, it can be cheaper to scale CPUs
than cryptocards.

In the more general crypto arena, if you're after HSM funtionality and
protected key storage, as much or more so than CPU-saving, then you're
going to be using such a device regardless.

> Unfortunately I don't have a PadLock CPU. I'm curious about  the performance 
> gains that would provide with monotone. I use my (pre-PadLock) Mini-ITX as 
> monotone server among other things.

I don't have any experience with this one.  If it implements some
extra special-purpose cpu instructions (that the compiler, or an
asm-coded library, would emit) to run instead of less efficient
general-purpose integer ops, or even if it's just cache-coherent with
the cpu, then it could be a different story to the above.

--
Dan.

pgp2aaOkwo5sZ.pgp
Description: PGP signature