> >> Someone had mentioned how fast Linux/Unix can search the hard drive
> >> for a text string.
> >> Well, that's a pretty simple and straight forward process that was
> >> optimized long before Linux and NT were around. And since I happen to
> >> have to make a search, why not benchmark it?
> >> So here goes.
> >> I can search for "www.ivillage.com"....
> >> through 6511 files...
> >> totaling 337,055,891 bytes...
> >> in 310 seconds.
> >> That works out to just over 1.087 MBps.
> >> The machine is a 6x86MX PR200+ with 64MB RAM running NT with SP3 and
> >> IE4.
> >> Okay Linux dudes...what have you got?
> >You need to check your math. 337,055,891 bytes/310 seconds = 1.037MB/s.
> No, we should just decide on what we mean by MB. How about this.
> 337,055,891bytes / 310seconds = 1087277bytes per second.
> >For comparison, i just grep'd through 819,091,456 bytes in my source
> >tree in 745 seconds (=> 1.049MB/s). That's on a PPro200, 128MB RAM.
> >Oh, and it also had 83 processes running at the time, including two
> >open telnet sessions, a remote X session, several server daemons...
> >and the real kicker: a similar PPro200 box running NT and building
> >from the same Samba-exported source tree that I was grep'ing locally.
> >I'd dare bet that's a fair sight more load (and particularly filesystem
> >load) than your NT box was seeing, and Linux still comes out ahead.
> I run a proxy server, and, at the time, NT was tracking about 28
> processes. But that doesn't mean anything if it's all sitting idle.
> Frankly, I'm surprised that a Pentium Pro was barely faster. The
> processor itself is much faster than my Cyrix.
Guess again. This kind of text search is almost completely I/O
bound. If you watch the CPU load during such a search, you'll
find that the CPU is more than 90% idle throughout the run. In
such a case, processor benchmark speed (particularly on such
things as floating point operations, where the PPro far exceeds
your Cyrix) means very little above a certain low minimum.
What's more important is clock speed (particularly on the system
bus, where you may well be ahead) and, even more so, disk speed.
Thus my running of another fairly disk intensive operation on
the same filesystem while performing your text search should
have been a significant penalty.
These comparisons are pretty meaningless because of the wide
variations possible in disk subsystems, but they certainly don't
give you reason to claim NT performs better than Linux.
Quote:> I don't know what the load was from your other processes so as far as
> I'm concerned Linux was slower. Shut down all your processes and try
> it again so you can get an accurate number.
I did. The numbers are still as meaningless as before, but they
are about 10% faster with little more than the single search
running. Of course, if you want to see the numbers jump, scale
down to a 486 and scale up to a cacheing RAID 5 subsystem; but
that will hardly prove that whichever OS you choose to run on
such a rig is faster than any other OS running a similar search
on a 1GHz Alpha 21264 attached to nothing but a tape drive.
--
Mark Hamstra
Bentley Systems, Inc.
by