Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 17 Feb 2000 21:52:51 +0100
From:      Brad Knowles <blk@skynet.be>
To:        Tom <tom@uniserve.com>
Cc:        freebsd-stable@FreeBSD.ORG
Subject:   Re: Initial performance testing w/ postmark & softupdates...
Message-ID:  <v0422082db4d20b7b0d63@[195.238.1.121]>
In-Reply-To: <Pine.BSF.4.05.10002171205530.1712-100000@shell.uniserve.ca>
References:  <Pine.BSF.4.05.10002171205530.1712-100000@shell.uniserve.ca>

next in thread | previous in thread | raw e-mail | index | archive | help
At 12:27 PM -0800 2000/2/17, Tom wrote:

>    Not really.  You could just use async updates instead of softupdates.
>  Or an OS that uses async updates.  Write caching metadata is always faster
>  than re-ordering it intelligently.

	We're getting several steps away from the primary reason I 
originally decided to run these benchmarks -- to get an idea of what 
kind of performance increase you might be able to expect if you got 
rid of synchronous meta-data operations altogether.

	Using async writes is not something I will be testing (or am 
interested in testing) except on OSes that do it by default.  In 
fact, so far as I know, Linux is the only OS that does this by 
default, and I can't think of too many OSes that are capable of doing 
it at all.

>    Perhaps.  Or it was the rather new SmartRAID V driver.  Or it was the
>  card (there are many different SmartRAID V) models.  Or the array was
>  optimized for sequentional io, instead of random io, not that RAID5 is
>  ever optimal for random io to begin with.

	Perhaps.  That machine still isn't being used for much, so I 
could fire off another test that used the internal drive, and see 
what that gets me.  I'd be willing to bet that softupdates still 
beats Linux ext2fs+async.

>    The rates reported by postmark are aggregates.  I've noticed that
>  actually throughput over the tests to vary by a 100% or more during the
>  tests.

	They may vary by 100%, but as an average, these numbers should 
not be too high nor too low, and the peaks shouldn't be all that much 
above nor should the valleys be all that much below.

	However, let's assume that the average is low, and that we should 
increase it by 100% -- double it, in other words.


	Well, if we double the numbers they got, 800KBps read and 1.2MBps 
write still isn't anywhere close to 12.5MBps available.


	Unless you're telling me that the inherent problem with NFS is 
that it can't sustain anything more than 1/20th of the available 
bandwidth due to protocol and other network issues, and if that's the 
case then things are *far* worse for NetApp and the other NFS vendors 
than even I think (used to think?).

	If that's the case, then as far as I'm concerned we can stop the 
conversation right here, since I know damn good and well that I can 
get ~18MBps on a four-way vinum striped volume with 10kRPM disks on a 
AIC-7890 controller, because I did it.  And even this isn't enough 
for the kinds of applications I have in mind.

>          Besides, latency of 100BaseT is rather poor too.

	Latency, now that's something I won't try to argue, especially 
when it comes to really chatty protocols like NFS.  However, I see 
this as just one more strike against NetApp (and NFS in general), and 
not something to be considered in their favour.

>    Not just their (NetApp) performance.  Disks and controllers in general
>  are a lot faster than they were two years ago.

	As I recall, the Adaptec 2940U2W and AIC-7890 controllers are 
pretty old themselves (the 3950U2 came after it, and now we have the 
Ultra3 controllers).


	In addition, this disk isn't coming anywhere close to what this 
controller is capable of -- the Quantum Atlas IV drive is several 
years old, old enough that I almost got one years ago for my PCI 
PowerMac 7200/90, which was the second generation of PowerMacintosh 
computers (we're now in the fifth generation).

	Even then, it wasn't the fastest around -- we have some ancient 
4GB Seagate "Cheetah" drives that are in Sun Ultra 1 and 2 servers 
that we are pulling out of production because the machines are just 
too old and creaky (Solaris 2.5.1?  Give me a break...).


	The machine I'm using for the tests is nowhere near 
top-of-the-line, either.  Again, it might have been top-of-the-line 
two years ago, but then NetApp was building "screamers" on Alpha EV4 
hardware, and with that kind of power, it should run rings around my 
older hardware.


	Let's not talk about old unless we consider that I'm using some 
pretty old stuff myself.

>                                                  Please forget all those
>  old results.  My point here is not to defend NetApp, but to say that old
>  results are meaningless.

	I could have put together a system like this two years ago, and 
then the results would have been immediately comparable.  If they 
would have been immediately comparable then, I submit that they are 
immediately comparable now -- with the exception of the software 
improvements that have occurred in the last two years.

>                            Big deal.  You managed to exceed some results
>  published over two years ago.  I bet my TV can exceed the postmark scores
>  you published today, two years from now.

	Perhaps.  But take a top-of-the-line TV that exists today, and 
try to exceed the scores I published.

	Heck, wait two years and take a TV that was top-of-the-line two 
years ago, and try to exceed the results I published.

	I dare you.  ;-)

-- 
   These are my opinions and should not be taken as official Skynet policy
  _________________________________________________________________________
|o| Brad Knowles, <blk@skynet.be>                 Belgacom Skynet NV/SA |o|
|o| Systems Architect, Mail/News/FTP/Proxy Admin  Rue Col. Bourg, 124   |o|
|o| Phone/Fax: +32-2-706.13.11/726.93.11          B-1140 Brussels       |o|
|o| http://www.skynet.be                          Belgium               |o|
\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/
     Unix is like a wigwam -- no Gates, no Windows, and an Apache inside.
      Unix is very user-friendly.  It's just picky who its friends are.


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-stable" in the body of the message




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?v0422082db4d20b7b0d63>