Hi Marcel,
Very interesting! I think we should discuss this the next SIMAP/CACE
meeting in Aarhus. We want, of course, that any application runs as
fast as possible, so we need to think about how we get there in a
systematic way. For instance, you call what you did a hack - is there
a more "official" way to do it?
regards, Ivan
Quoting Marcel Keller <mkel...@cs.au.dk>:
Hi friends of VIFF,
I've now implemented the hack I mentioned at the last SIMAP/CACE
meeting. The quintessence of the hack is that we do network
communication every time a multiplication or open operation is
scheduled, not only after returning the control back to Twisted.
As you can see in the attached graphs, the results are much better
and more accurate than without the hack. The graph also shows that
inversion by masking still is slower than inversion by
exponentiation. However, the gap is smaller because I optimized PRSS
which is used more with masking than with exponentiation. I guess
that masking is slower because it needs more local computation and
the network latency between the test machines is quite low (about
0.1 ms), which increases the impact of local computations.
If you want to try out the hack yourself, you find two patches
attached. One is against Twisted 8.2.0 and one against the current
tip of my VIFF repository:
http://hg.viff.dk/mkeller/rev/d522f9b14b49. The PRSS optimizations
can also be found there.
Furthermore, a new version of the document describing the AES
implementation is attached. It corrects various errors of the
previous version.
Best regards,
Marcel
_______________________________________________
viff-devel mailing list (http://viff.dk/)
viff-devel@viff.dk
http://lists.viff.dk/listinfo.cgi/viff-devel-viff.dk