Both blogs were discussing how ZFS is great for a mail server workload as benchmarked by Postmark. I have great respect for Sun’s performance team with alumni such as Adrian Cockroft [now with Netflix] and current superstars in their team such as Jim Mauro and Richard McDougall and I think ZFS is an innovative filesystem which eases a lot of pain
Thus it pains me a lot that they are trying to pass of a benchmark (Postmark) which does not have a single fsync(2) as appropiate for a mail server. A mail server has a number of crtical places where it has to do fsync in order to reliably write the message to queue. Failure to do so means that MTA has absolved responsbility.
To me a benchmark which models ISP mail workload correctly is Bruce Guenter’s benchmark . I think Sun should invest in porting this over to the filebench framework. This would command more respect than trying to pass of Postmark as representative of a ‘mail server’
Colm “Punish thy Server” MacCárthaigh has published his awesome benchmarks with the Sun T2000. Colm is also looking at ssl-offload benchmarks for the T2000. I hope my engine(3) patch for flood gets to see some action.
As Chuck Norris fans would say, the Niagara roundhouse kicked the Itanium and the Xeon.
Patch compiles on my AMD64 box, now I just need to scrounge for a Niagara so I can see if it takes advantage of the insane RSA performance to be the “Speedy Gonzales” of https load testers
A shout-out to anyone in Sun. Help me make the Niagara look good
Sun has recently released it’s new UltraSparc T1 based systems (aka “Niagara”) boxes. These boxes are basically single CPU boxes but the CPU have 8 cores with each core having 4 hardware threads on them. Solaris 10 sees 32 CPU’s on this box. Workloads which are threaded work very well on these boxes.
The other thing cool about the Sun Niagara is that they have phenomenal RSA performance which seems to be accessible via the SSLCryptoDevice directive to Apache. With Apache 2.2 mod_proxy showing a lot of improvement, this would make Apache with the worker mpm (small number of processes with lots of threads per process) combined with mod_ssl an exciting combination to run.
It seems a bit strange though that Sun is recommending that people compile Apache with the prefork mpm. I guess that might be appropiate if Apache were to be compiled with PHP or some other module but I would expect that for an HTTPS offload workload, then worker might scale better.
The other interesting bit would be to modify Apache Flood to have support for engine(3). This would allow for a very fast threaded ssl aware http benchmark which would take advantage of the RSA speedups within Niagara.
It looks like some Apache committers might be getting their hands on this box soon.
Now, if only there was support for SNI via mod_ssl or if mod_gnutls support engine(3) then these boxes are likely to be no-brainer for ISP’s to host SSL frontends.
Today the cost of SSL enabled websites increases due to
- cost of doing SSL computations
- The requirement of one IP per hostname hosting SSL
With increasing CPU performance (particularly the AMD Opteron) which totally rule in terms of RSA crypto performance, point (1) is slowly becoming a non-issue. point (b) is still an issue
point (2) is being addressed via Server Name Indication.
which is currently only supported in Opera 8.0.
IE 7/Vista will also support SNI
Guess which company the engineer works who has the bug assigned to him. You are right, It’s Sun Microsystems.
So here’s a shout-out to the Sun bloggers out there. If you believe that making it easier for webhosters (who may purchase Sun hardware if they find that SSL performance on the Sun Niagara boxes screams to easily host virtual SSL hosts on a single IP is a worthwhile proposition, then I encourage you to evangalize within the organisation to enable the engineering resources within Sun so that Mozilla/Firefox have support for SNI at the earliest.
Remember that a large percentage of the world isn’t going to move to Vista so SNI support in Firefox may even lead to a faster adoption of the browser and help in standard adoption
Well, I can think of one class of deals which Jonathan can easily win but isn’t getting into today. These are the Linux/FreeBSD NAS boxes which use 3ware IDE Raid controllers. Just do a google search or search Redhat/Suse’s bugzilla to see the pain people face monkeying with NFS and the linux filesystem of the week.
With ZFS and Sun’s robust NFS stack (Check out which company has a lot of slots in the upcoming Connectathon), Solaris can pretty much own the market in the low-cost NAS box. Track a few mailing lists and sales staff can cold-call the appropiate sysadmin who has pretty much detailed his pain points when trying to setup a robust NFS server. Maybe there are cross-sell opportunities for StorageTek products.
Jonathan, all this needs is the driver team in Beijing and/or your IHV/ISV engagement teams to get cracking with 3ware at the earliest.
Sun currently uses the tag line “ZFS: The last word in filesystems”. From what I read and saw, I immediately felt that sysadmins who start their career with zfs will never understand the pain felt by people who’ve gone through volume management, fsck, moving filesystems around. A lot of sysadmins have missed crucial moments with their loved ones because they were fighting fires. ZFS makes that a bad dream
I feel Sun Marketing should market ZFS as
“ZFS: Because your family deserves it”
Congratulations to team ZFS and particularly to Jeff Bonwick who had the vision and courage to “think different”
A shout out to our friends at Sun Beijing, there are a truckload of installations running Linux/BSD boxes using 3ware. Get a 3ware/Areca/Qstor driver in Solaris 10 at the earliest and Solaris/ZFS becomes the NAS OS of choice.
I cut my teeth with SunOS/Solaris. With the recent release of Studio 11, it’s going to be very compelling. I hope the Solaris hackers can put some love into the installer and provide some update tools like yum,apt-get.