Slashdot Mirror
Sun Super Computer May Hit 2 Petaflops
Fletcher writes to tell us that Sun Microsystems has revealed their "Constellation System", a new supercomputing platform that the company hopes will put them back in the running for top dog in the supercomputer race. "The linchpin in the system is the switch, the piece of hardware that conducts traffic between the servers, memory and data storage. Code-named Magnum, the switch comes with 3,456 ports, a larger-than-normal number that frees up data pathways inside these powerful computers. 'We are looking at a factor-of-three improvement over the current best system at an equal number of nodes," said Andy Bechtolsheim, chief architect and senior vice president of the systems group at Sun. "We have been somewhat absent in the supercomputer market in the last few years.'"
IBM Blue Gene/P update slated to run at 3 petaflops.
Exactly what markets has Sun been going gangbusters in since about 1999?
why the Java market, it's everywhere. of course, Sun hasn't made a thin dime off of it, but the market sure went for it.
If you RTFA you will note that, actually, this particular system is built around the Barcelona architecture (from AMD). It remains to be seen if T2 and later on Rock will really be competative against AMD and Intel.
]{
Actually 3456 is 12 X 12 X 12 X 2. It's not actually a 3456 port router, it's a fat tree of 24-port router modules. Each rank 1 & 2 module has 12 ports down and 12 ports up. The rank 3 modules have 12 ports down, and 12 sidelink ports to one another. Thus you end up with a 3456 port, rank 3.5 fat tree all in one box.
The latest top 500 list is out today. http://www.top500.org/stats/list/29/osfam/ is the link to the OS statistics. It appears that Tamagochi farms aren't that popular amongst the high performance computing crowd.
I'll add:
This is not an unusual arrangement for existing infinaband networks. The distinction is that they have all of these 864 switch modules in a cabinet, and the wiring is probably traces on a backplane, rather than flexible cables. This improves the reliability, reduces the cost, and makes it a whole lot easier to install. That may sound silly, but you're talking about 10,000 cables, each with endpoint connectors on each end. Even buying in bulk, that's a lot of money in cables.
> in the PC world we're finally seeing this architecture recognized as new Intel chips tout their front-side bus and cache
> more than sheer increase in speed.
> This SUN machine is a bigger-scale example of the same.
No, not really - parallelism has of course been around forever. But its application for high performance computing has been constantly demonstrated on large servers for the past 15 years. This is back when parallelism on intel hardware might at most have meant two cpus. And that was rare.
MPP (Massively Parallel Processors) systems like Teradata and IBM's SP2 (aka DeepBlue - that defeated Kasparov at chess) successfully demonstrated great performance for the dollar back around 1994-1995 or so. These were originally designed around data mining and math computations - but found most of their sales in data warehousing. Meanwhile, CRAY was complaining that not all problems were good candidates for this kind of more cost-efficient hardware.
By 1998 you could put db2 or informix on a hundred-node SP2, each node consisting of an eight-way SMP, each with its own dedicated storage. Queries on that old system were lightning fast compared to most other options. I worked directly on SP2s and worked with a team that has a 128-node one. Oracle & Sun eventually ecliped these solutions with massive SMPs. But much of that was more due to Informix's financial issues than technical merit - since Informix and DB2 (and of course Teradata) on MPPSs easily out-scaled oracle on SMPs. The SMPs were easier to adapt to application design changes, but the MPPs were easier to grow indefinitely large.
These newer solutions are just more of the same thing - you've still got the same challenges in:
- tons of OS and application images that must be consistent
- node communication bandwidth (major selling feature of all these solutions are proprietary internal networks)
- failover (how do nodes failover, especially if they have any dedicated resources)
- scheduling (how do jobs get assigned to nodes)
So, they're much bigger and faster than 10-15 years ago - and I'm sure there's got to be some cool innovation going on under the hood. But nothing looks fundamentally different from then. And nothing here has been inherited from the pc world.
...but will it run linux?
The first one of these being built, at the Texas Advanced Computing Center, is in fact running Linux, not Solaris (See this Register article). Sun will support both.
Yep. We are implementing a datawarehouse app and after seeing the poor performance we are getting out of Oracle on Windows for our ERP product we are researching different solutions. Right now Sun is in the lead. They are about 20% less than our Windows solution for the overall solution and we expect it to be more scalable as we will only be partially populating the DB server out of the gate. They claim they can do with 4 middle tier boxes what our Windows solution provider has speced 13 Windows boxes for. I can't wait to see the results of the bakeoff. I'm primarily a Windows/Citrix guy but I've admin'd Solaris and Linux in past and I won't ming keeping my Solaris skills up to date if it's the best solution =)
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.