[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: MD5 performance
> MD5 Performance update:
>
> On the 64-bit RISC processor that got 20 Mbps using the
> RFC-1321 code,
I observed the same figure on 100MHz HP (32bit RISC).
As MD5 is 32bit based, 64-bitness does not help.
> there is new news that by hand tweaking the code in a
> one can get about triple the speed.
In general, "hand tweaking" can not do better than compilers on
RISC.
It is obvious from my "prof" command analysis that the two thirds of
the time is wasted for bytewise-I/O and byte-aligning (byte-aligned
IPng is really a good news), removing which will results in triple
performance, 60Mbps.
> I'm being deliberately vague here on exact performance because
> slightly different system configurations were used by different groups
> to get these numbers.
Counting the number of basic operations (about 700 for 64 bytes),
theoretical maximum for software approach is at about 1 MB/sec for
10 MIPS.
Considering that super scaler boosting does not help so much for MD5,
maximum speed with 100MHz processor is at about 80Mbps and we are now
close to that limit.
> My concerns over MD5 performance are significantly abated by these
> more recent results. It looks like IP/FDDI, IP/HyperChannel, IP/ATM,
> and IP/PPP/SONET may have performance problems with MD5. Most users
> are at Ethernet speed or below and will be fine.
We will soonly have a new news that most users are much above at
Ethernet speed.
> IMHO, we need to
> consider selectively authenticating packets in any event
Or hardware MD5 generator.
Masataka Ohta
References: