Although the portions of the systems subject to Moore's law helps
there are critical components which have not improved much at all in
the same time period.
Can't disagree with that.
For example, disk speed isn't all that much faster today than it was
10 years ago. Where CPUs have improved by about 1000-fold, disk has
improved maybe 10-fold or so, and mechanically much less (1993,
3600RPM, 2003, 7200-10,000RPM, seek times have barely changed at all
remaining in the single-digit ms range, albeit cheaper.)
I don't think this bit is quite accurate.
Lets look at discs: First of all, the common discs in 1993 were 2400 rpm,
not 3600 - if think of 3600rmp and 11 ms seek time for 1993 you had better
compare it with 15k rpm and 4ms seek time for 2003, alternatively you can
talk about 2400rpm and 15ms for 1993 and compare with 10krpm and 8ms for
2003. Second, in 2003 drives have substantial internal buffering and do
plenty of read-ahead; for a remailer/list exploder application this reduces
the number of seeks quite a bit and also reduces the impact of rotational
latency - the application is not a random access one, where perhaps 12 times
would be a reasonable disc speed up estimate, but one where the internal
drive buffering and read-ahead has a substantial effect so that there has
been a much greater improvement. Third, head settling times are much lower
so that short seeks are now zero cost - in 1993 even a two track seek would
lose you a rev. Fourth, surface density is much higher, so that transfer
duration is much lower, and we have much larger mainstore so that we can
cache in the OS much more effectively store. Fifthly, with this particular
application I can use two discs in such a manner that I just about eliminate
seeks of more that 1 cylinder, and 1 cylinder seeks are zero cost on modern
discs.
Now lets look at processors: do we have anything like a 1000 times speed up?
Well, clock times do have a 1000 times speed up but I've never had much
faith in clock time as a measure of speed. How much has FSB bandwidth
changed? Or main store bandwidth? - for some applications that will be a
much better measure. So, unless your crunching numbers that fit in L1 cache
you have nothing like a 1000 times speed up.
So, unless you're crunching numbers which all fit in memory, Moore's
law is not a reasonable gauge of e-mail processing improvement. And
often this sort of processing can only proceed as fast as its weakest
link, which is likely to be disk speed.
Much more likely to be main store bandwidth for the application in question,
I think, but thet too is obscured by caching and of course it depends on bus
width (but most systems use a store bus narrower than a complete L3 cache
line, so it's not as much obscured as it might be). Of course that doesn't
affect the conclusion - Moores law is not a useful gauge of e-mail
processing improvement.
Cheaper RAM enabling much more aggressive and accurate caching of
course clouds this picture somewhat.
It certainly does. I think it clouds it enough so that the bottleneck will
not be where you think it is.
But to imply that because CPUs have gone from 1MHZ to 3GHZ that we can
now process 3000x as much mail is unsupportable and misleading.
Here we are in total agreement.
Tom Thomson
_______________________________________________
Asrg mailing list
Asrg(_at_)ietf(_dot_)org
https://www1.ietf.org/mailman/listinfo/asrg