The Power Systems JS12 and JS22 Blades Versus Other i Boxes
July 14, 2008 Timothy Prickett Morgan
Over the past several weeks, I have given you the detailed feeds and speeds of the new Power6-based Power Systems machines running i5/OS V5R4 and i 6.1 and shown you how the new entry and midrange machines compared to the Power5 and Power5+ generations of System i gear. This week, I will bring it all back together and throw the JS12 and JS22 blade servers into the mix–the first blades from IBM that can support an i operating system–inside the new BladeCenter S chassis for small businesses.
Big Blue may have brought blade servers to the OS/400 and i5/OS platform a little late in the game–six years behind Windows and Linux blades and five years behind PowerPC blades that could run AIX or Linux–but by all indications, the company is intent on making BladeCenters sporting Power6 blades and virtualized I/O a credible economic alternative to rack-based and tower-based Power Systems 520 and 550 servers that are in roughly the same power class. No kidding. I checked my math in the comparison sheet I made to examine the performance and pricing of JS12 and JS22 blades compared to Power5+ and Power6 System i and Power Systems machines three times. I didn’t believe it either.
IBM’s first foray into the i-based blade space, the JS22, did not get off to a particularly good start last November when this quad-core Power6 blade was announced. For one thing, IBM charged a lot of money for the configured blade and operating system, which wasn’t even available until the following March, and for another, it put the machine in the P20 software tier, making it wickedly out of reach for a lot of small and medium businesses that have software in the P05 tier and start having heartburn when they see a machine in the P10 class; the P20 tier gives many a heart attack, and actually prevents them from buying the iron at all. As I said last fall, IBM needed to get a smaller JS12 blade with one or two cores–four cores and 13,800 CPWs is far too much oomph for SMB shops, at least for the OS/400 and i5/OS workloads–and a P05 tier out the door, and lo and behold, IBM actually listened to the community. (I was not the only one saying this needed to be done, of course.) And, to put some icing on the cake for a change, when the spring of 2008 rolled around and IBM delivered the Power Systems lineup with this JS12 blade and the i 6.1 operating system, the company went back and put the JS22 blade in the P10 software tier. Where it belonged from the getgo.
And because of the aggressive pricing of the BladeCenter S i Edition, which bundles the blade chassis, a JS12 blade, switch and disk features, and the i 6.1 platform (with the required Virtual I/O Server for virtualizing I/O on the blades as well as the PowerVM Standard Edition for setting up logical partitions) all together for a discounted and, at least compared to Power5+ and Power6 rack and tower servers, very reasonable price. This is, of course, exactly what IBM planned–an agnostic marketing position between tower, rack, and blade servers delivering the same i 6.1 platform and roughly the same raw online transaction processing performance. I say roughly the same raw OLTP performance because I am not yet convinced that the JS12 and JS22 blades are not I/O constrained and that those CPW ratings of 7,100 for a dual-core JS12 and 13,800 CPWs for a quad-core JS22 running i 6.1 are a lot more theoretical than practical in a data center. I am not saying that customers cannot realize this relative performance, but you sure can’t do it with a blade server with one or two disk drives on the blade. No way.
That’s why in the configurations I created, I put one or two SAS blades (really just disk enclosures that fit into the chassis) with six disks each and used RAID 1 mirroring data protection to make something close to a comparison to the RAID 5 disk arrays I configured in the System i 520, 515, 525, and 550 servers and the Power Systems 520 and 550 machines I have looked at in prior comparisons. The table accompanying this analysis brings all of those machines into one giant table. By doing this, I have put a similar number of disk arms on the machines as I put into the tower servers. Because of RAID 1 mirroring, you have to cut the number of disk arms for the workload in half, since they are paired up. I am also assuming that there is a way to make use of the disk drives on the blades themselves to host the i 6.1 operating system and the VIOS partition for virtualizing I/O, but this may not be the case. What I am trying to avoid doing in these comparisons is resorting to an external DS4000 disk array and Fibre Channel links. This is a perfectly reasonable option for larger midrange shops as well as enterprise customers using the i platform, but it is overkill for SMB shops that are used to–and that prefer–internal disk arrays.
Like last year’s System i 515 and 525 servers and this year’s Power Systems 520 M15 and M25 machines, the JS12 and JS22 blades have their i 6.1 licenses priced based on a base licensing fee and then per-user fees on top of that. The System i 550 and Power Systems 550 boxes are not priced based on users, but rather assume unlimited users and have per-processor fees for either i5/OS V5R4 or i 6.1.
To help make your capacity planning a bit easier, I calculated the cost per transaction per minute for the machines based on the CPW benchmark and my conversion of that rating to estimated TPC-C online transaction processing performance. (This lays the groundwork for comparisons of these Power6 machines to AIX and Linux running on Power as well as to other Windows, Linux, and Unix platforms.) I have also added users for the earlier machines that did not have them; I want you to be able to compare the price, performance, and price/performance of machines based both on OLTP throughput and end user seats. I don’t think of the number of users I put on the 550-class boxes or Power5+ System i machines that were not user-priced are untypical; if anything, they are probably heavily loaded, since systems are usually bought based on their peak batch capability as much as on OLTP performance.
In the comparisons, I also removed Software Maintenance fees from all of the systems, even though IBM requires customers to buy it, because I think customers should have the choice of maintenance services or not (or IBM or otherwise) and because I want to isolate system costs from support costs. I also configured a the JS12 with one fewer core than the BladeCenter S i Express Edition has and priced the JS22 setup as if IBM offered a BladeCenter S i Express Edition for this blade, which is does not. I think anyone looking for JS22 blades should insist on paying the same $1,946 for the S chassis, the same $159 for the CD drive, the same $795 for the storage module for SAS drives, the same $999 for the SAS switch module, the same $699 for the copper pass-thru module, the same $79 for the pass through cable, and the same $10 for the power cord. Period. From there, you add blades and storage and other stuff.
One more thing that I did not make clear in the comparisons thus far. The improvement in price/performance is based on making the best comparisons for successive generations of iron. So, for instance, where there were user-priced 515 and 525 machines in the line, the new Power6 M15 and M25 gear has their bang for the buck reckoned against these 515 and 525 machines, not the earlier i5 520 Standard Edition and Enterprise Edition machines. If there were not user-priced machines (as was the case with the 550-class machines), the newer Power 550 and JS22 are compared to the Power5+ 550 box. The JS12 and JS22 are not being compared to their Power 520 and Power 550 brethren directly, but relative to a common, earlier machine–in the case of the Power 520 and JS12, that would be the System 515 and 525.
As you can see from the table, the blade servers running i 6.1 hold their own against the rack and tower Power6 machinery in the same power class. That is true even after loading up the memory and disks and a SAS-based LTO-3 tape drive, which costs $3,250. (The rack and tower machines all have a 36 GB 4mm tape drive, which costs a lot less, configured in them, which is standard.) In many cases, the blade implementation is a lot cheaper, even if customers will have to sacrifice some performance. The reason for this is simple: IBM is charging relatively high prices for Power6 processors and i 6.1 on the rack and tower machines, and it is charging a lot, lot less for the same exact components on the blades. How low, you ask? Well, a JS12 blade costs $200, a dual-core processor for it costs $3,140, a processor core activation costs $100; 2GB of memory costs a mere $219 and a SAS disk for the BladeCenter S chassis costs only $429. PowerVM Standard Edition is a free on the bundle, and i 6.1 costs only $,1795 per core. As far as I can tell, there is no Enterprise Enablement fee on these blades–you want 5250 capacity, you just use it, just like with the 515 and 525 machines from last year and the M15 and M25 machines from this year. Not so on the System i 550 or the Power 550, where you have to pay $40,000 per core for i5/OS or i 6.1 and $50,000 per core for 5250 enablement. Ouch!
So, from the raw numbers, the JS22 is a wickedly better choice for supporting i 6.1 applications than a Power 550 M50 machine–and mostly because of the high processor prices, higher core activation fees, higher i 6.1 licensing and 5250 enablement fees. That’s even with IBM charging $10,995 per core for i 6.1. You can even toss in the $50,000 fee to turn the JS22 into an unlimited user box (akin to the M50) and it doesn’t even dent the comparison all that much. (I did not do that in my comparisons.) The JS22 blade is so much a better deal–particularly when you consider that the Power 550 is still in the P20 software tier–that I wonder what the problem is. I don’t know this for sure, but IBM’s own feeds and speeds, show the JS12 and JS22 sporting a lot less memory and I/O bandwidth than their Power 520 and 550 counterparts. And on OLTP workloads, which is the bread and butter of the AS/400 base, this is a big deal.
A spec sheet I obtained from IBM shows the JS12 and JS22 having 21.3 GB/sec of memory-to-processor bandwidth and 5.8 GB/sec of I/O subsystem bandwidth, compared to 32 GB/sec of memory and 14 GB/sec of I/O bandwidth for the Power 520 and 128 GB/sec memory and 14 GB/sec of I/O bandwidth for the Power 550. This, my friends, could be the catch in all that attractive pricing. The other gotchas with the blades are that the internal SAS modules for the BladeCenter S chassis do not support RAID 5 data protection–you’ll have to use an external DS disk array from IBM to get more capacity and disk drives–and that the configurations I showed with a dozen SAS drives (mirrored, so it is really only six drives’ worth of capacity) are the maximums for the chassis. Moreover, the blades do not have the 32 MB of L3 cache per Power6 chip, which is shared for the two cores and which helps boost performance and presumably memory bandwidth significantly. And finally, the JS22 peters out at 32 GB of memory, while even the JS12 can do 64 GB. The Power 520 can do 64 GB and the Power 550 can do 256 GB.
So test your workloads before you buy and plan your memory configurations and disk arrays carefully. And if you are going with blades, tell me what you did and how it worked out.