Posted Wednesday 7th November 2007 22:39 GMT
I can see two heatsinks. Does this mean that a pair of processors share a heatsink, or does "four processors" mean two dual core sockets? I'm a Sun man, and Sun use the term "dual core processor" to mean two cores in one socket.
Perhaps the terms "socket", "core" and "thread" are suitably unambiguous. I'd be interested to discover how these blades will carve up into LPARs: with SPARC LDOMs, we can assign down to the thread level into an LDOM, then carve out processor sets in each Solaris instance ...
Q: "How many processors does my server have?"
A1: "Well, that depends on what you mean by processors, and what you mean by server".
A2: "You're paying for capacity, not processors or a server".
Are IBM expecting these blades to be used as a single instance, or to be carved up into virtualised server capacity?
Posted Thursday 8th November 2007 01:11 GMT
I want one. No, I want a rack of them. Of course I don't have anything that needs it, but that's besides the point. I want to be in the Top500 list. Me.
Oh, well, everyone can have dreams.
Seriously, I looked at the Toop500 list to see the systems that had the fewest number of processors. What was there? Hitachi with Power5+ processors. And now here is Power6. Oo la la!
Posted Thursday 8th November 2007 01:21 GMT
Power6 CPUs are dual core parts - the article is written in a slightly misleading way.
Assuming your application runs well on PPC this should be pretty far up the list of bladeservers, speed wise.
Definately interesting - although personally I plan to wait a little and let them fix any issues before thinking about buying any.
Posted Thursday 8th November 2007 06:09 GMT
Two POWER6 chips - two heatsinks - two cores per chip - 4 cores per system.
And each core can have up to 10 partitions, so a single blade can have up to 40 micropartitions (SPARC LDOMs are their version of IBM's micropartitions). Here, instead of partitioning based on number of threads, you partition based on capacity requirements, 10% of a core being the minimum, and going up in 1% increments. This feature is called Advanced POWER Virtualization, and it is available on POWER5 systems and the PPC970 blade (JS21) as well.
Posted Thursday 8th November 2007 11:44 GMT
I've spent the last few days being blue-rinsed by IBM on p6 & AIX 6. Very nice... As well as the v. desirable new hardware, AIX 6 (nee 5.4 :-) has got some damn clever tricks too.
IBM's LPARs let you carve your machine up as you see fit (starting from POWER4). POWER5 brought in the micro partitions (see Cesar's comment), which let's you finesse things considerably.
AIX 6 brings in Workload Partitions (WPARs), which let you effectively slice your O/S instance even further (I think this is Sun's containers style operation?). Because it's AIX 6 doing this, the function is available on POWER 4, 5 and 6 (but not older kit) - you *don't* have to have p6.
Maybe a good way to tell how this plays is to see how the competitors (HP, Sun) try to play down IBM's announcements. Right now, I can't see any real compelling reasons not to play the IBM game, but then I may be a bit biased (no, I don't work *for* IBM, but I've spent a whole lot of time working *with* IBM - anybody remember 6150 / RTPC?).
Posted Thursday 8th November 2007 18:51 GMT
It's quite sweet. I installed/configured/set up our application for a customer on a 9117-MMA (570 POWER 6 with 16 cores) a couple months ago. I made a partition to load up the beta AIX 6 during testing and I liked what I saw.
Those chips run stuff nice and fast. I can’t wait until they get rolled out on the other product lines.
Posted Thursday 8th November 2007 21:58 GMT
The POWER6 processors are very powerful. But IBM's BladeCenter blades were designed in the era of single-core x86 processors. This limits memory capacity and I/O bandwidth.
The POWER6 p570 can support up to 96 GB RAM per dual-core processor. The POWER5 servers could support up to 64 GB RAM per dual-core processor. This blade only supports up to 16 GB RAM per dual-core processor. So memory resident apps need not apply.
But let's assume your data is not in memory. That means you will access it over the network. But unless you use the dual-port 10 Gigabit Ethernet card or a dual-port, double data rate InfiniBand card, the network will be a bottleneck on blades with this level of performance.
My guess is most POWER6 blades will be used in HPC clusters using InfiniBand or 10GE networks. But it will have to be applications which do not require large amounts of local RAM.
Right now the POWER6 blade looks like it will be very fast at waiting.
Posted Friday 9th November 2007 07:49 GMT
At first i won´t expect P6 performance of this blade. This baby has only 4 GB of cache. A real P6 has 40 GB. I assume, that this will lead to reduced performance. And in some document from IBM i found some hints that my assumption is not totally incorrect. I summarized my findings on http://www.c0t0d0s0.org/archives/3693-P6-Blade-or-The-effect-of-having-less-cache-on-a-Power6.html
But i have a different problem with the design. From my knowledge the IBM Bladecenter is a front to back design. In this case the airflow will be heaten up by this monster heat sinks and strive over the DIMMS and then over the single not-hot-swappable harddisk. The RAS story of this blade must be horrible ...
Posted Monday 19th November 2007 22:18 GMT
I guess the GB you speak off, are intended to be read as MB ;-) ........
yes, the blade version is not meant to do data-mining. not enough I/O slots to get to your data real-fast. But it wil do nicely in many cases.....
no, it's not a hot-chip.
something in the order of 60 W per processor (=core in IBM speak)
120 for a socket (=chip in IBM speak)
For the js22, it totals out to 376 watt maximum for a 2 socket, 4 core js22 blade with disks, memory etcetera ( memory takes a lot of power too, as a matter of fact).
you can get the details here: ( and a lot of other information on pseries as well.... )
http://www.ibm.com/common/ssi/fcgi-bin/ssialias?infotype=pm&subtype=br&pdf=yes&appname=STG_PS_USEN&htmlfid=PSB01628USEN
Back to the article