cerise at littlegreenmen.armory.com
Fri Jul 26 19:49:54 EDT 2002
Sergey A. Lipnevich said:
> My Dell Inspiron 5000 with Pentium III 700 MHz and 256
> MB RAM just showed a faster result for tempfs while
> building openldap :-). Now, I'm sure that disk-based
> compilation would be even slower if benchmarked
> properly because GKRellM was showing quite clearly how
> effective reiserfs cacheing is working on openldap
> source files. That means, if I rebooted before
> compiling off reiserfs, there would be additional time
> spent on reading sources from the disk. Anyway, I
> wouldn't proclaim tempfs losing before seriously
> testing it in terms of performance and fragmentation
> impact, before saying "shown by benchmarks" etc. I'd go
> as far as saying that nothing on this list has shown
> that anything is wrong with choosing tempfs.
> Take care!
I'm confused. You make it sound like you benchmarked both of them,
but your discussion about GKRellM suggests that you only tried it using
tempfs. If you haven't run time on both, then you haven't benchmarked
I have a fairly rigorous analysis which I sent to AFK some time ago
showing that tempfs will run in greater or equal time to the native
filesystem. It makes some assumptions (e.g. that disk access is the
only thing that needs to be considered, that caching works pretty well,
&c), but I don't think it hurts the argument. I suppose I could post
it to the list if there's demand. I doubt there is.
I'm a bit confused as to your discussion about rebooting as that
doesn't seem to have any reference to the issue (aside from good
benchmarking principles). At any rate, it's only sensible that
removing the files from the disk would remove any concerns about
caching unfairly favoring one run over another.
All of the numbers posted (excepting AFK's) have favored using the
native filesystem (ext2, ext3, and reiserfs have all been involved) on
disk. There's been ~5 or so of them. I'll accept that I am perhaps
assuming that the benchmarks are fair
In another couple of weeks, I'll be back on local console. I've
been planning to install Source Mage over my present Slackware
installation and I intend to make good on it. Give around a week or so
more of modifying the existing Sorcery to allow for usage of the native
filesystem and I'll post benchmarks along with the script used.
I'll grant that there have been no numbers on fragmentation (from
either side). However, given the numbers and the assumption that those
who benchmarked did so properly, I think the claim is absolutely fair.
The numbers have shown a speed decrease using tempfs. It's trivial
that there's an added memory component using tempfs. Given those two
facts, then it would take some pretty evil fragmentation in order to
offset the claims, especially given the magnitude of the difference in
More information about the SM-Discuss