IBM Launches Power7-Based Cloudy Stacks
October 18, 2010 Timothy Prickett Morgan
Last summer, when IBM launched its first CloudBurst virtualized private cloud infrastructure stacks, they were built on its BladeCenter blade servers and used the company’s Xeon-based HS22 blades. Last week, as Big Blue updated the X64 variants of the CloudBurst stacks, it rolled out the first versions of the cloudy infrastructure based on Power7 processors. But instead of using its PS700 or PS701 blade servers, IBM chose the workhorse Power 750 server as the building block.
The CloudBurst setups, you will remember, are a pre-integrated stack of virtualized systems that include servers, storage, switches, virtualization hypervisors, management tools, and a self-service portal for end users that are sold under a single product number with a single price. The idea is to do for private clouds installed behind the corporate firewall what supercomputing customers have been able to acquire through so-called “bright clusters” for the past decade. Companies don’t want piece parts, they want finished, usable gear and they want it at a discount off list price. IBM sells the CloudBurst setups in several different sizes, like men’s T-shirts, although it is drifting a bit on this with the Power7-based cloudy setups. (You know how hard it is for Big Blue to stick with a naming convention of be consistent.)
The big changes with the CloudBurst V2.1 setups on the X64 front include a move to IBM’s HS22V memory-dense blade servers and the latest six-core Xeon X5660 (2.8 GHz) processors. The HS22V blades come preconfigured with 72 GB of DDR3 main memory and are able to support 36 virtual machines, according to IBM’s sizing algorithms. (That works out to three VMs per core and 2 GB of real memory and 4 GB of virtual and overcommitted memory per VM, if you ignore overhead from the hypervisor, which you can’t but we have no idea how much overhead that is on either X64 or Power machines.) The blades are also configured with VMware‘s latest ESX Server 4.1 hypervisor, which came out in the summer and which has lots more bells and whistles.
There’s a stack of management software called CloudBurst that runs on a management server (a System x rack server in this case) and some more management software runs on one of the blades in each chassis (leaving you 13 blades for actual cloudy workloads) and a new self-service portal called Tivoli Service Delivery Manager. This latter bit of code overlaps with the CloudBurst software functionality, but the self-service portal allows for end users to request their own virtual machines and get them approved through a workflow process with the IT department.
Here are the current configurations for the HS22V blade clouds:
IBM’s Xeon-Based CloudBurst stacks. (Click graphic to enlarge.)
In that small configuration, IBM has rounded down the number of virtual machines to 100, and is charging $220,000 for the whole shebang. Yes, that sounds like a crazy amount of money for three blade servers running 100 virtual machines (108 actually by IBM’s own sizing), but it works out to only $2,200 per VM for the base configuration. An HS22V blade with two X5660 processors and 24 GB of memory has a list price of $5,435 using 4 GB memory sticks; boosting that to 72 GB increases the price by $11,280, so the blade costs a stunning $16,715. So it costs another $167,150 at list price to fill out the remaining empty 10 blade slots in the BladeCenter chassis, assuming there is no big discount, and that makes the cost of a 13-blade cloud $387,150. Who knows what IBM is charging for the VMware hypervisor, but it could be as low as $500 for the ESXi 4.1 embedded hypervisor or a couple grand per blade for the full ESX Server 4.1. IBM is also boosting storage to get to this larger configuration, and I would not be surprised if the price was in the range of $450,000 for the medium configuration. That works out to around $978 per VM, a much more sensible price.
So what about the Power7-based CloudBurst stacks? Instead of using its blade servers, which offer better density than rack servers, IBM has chosen its Power 750s, which are 4U rack servers with four processor sockets per box. As you know from reading the past several issues of The Four Hundred, the Power 750 is a lot more expensive per unit of performance than the PS700, PS701, and PS702 blade servers or the entry Power 710, 720, and 730 servers. The Power 750 therefore seems an odd choice for a company that is trying to show that its Power-based systems offer better virtualization than blades. After mulling it over, it may have something to do with I/O and memory bandwidth, or it may simply be that IBM has lots of Power 750s laying around all clustered up for PureScale clusters or HPC setups.
The Power 750 racks offer no advantage in terms of memory (128 GB per socket) or logical partitions (10 per core) compared to the single-socket PS701 blade server. IBM is putting 11 Power 750s across five racks to host 2,960 logical partitions–16 cores of the first Power 750 are used to run the CloudBurst management software) plus two Hardware Management Consoles to manage the PowerVM hypervisor. Although the Power 750 is currently topping out at 160 partitions per machine in a standard configuration, IBM says that it can host 280 partitions in the CloudBurst setups. (And based on the 10 partitions per core PowerVM can do, it should be 320 partitions). At any rate, to get to that 2,960 partitions, you have ten and a half Power 750s spread out across five racks running virtual AIX, Linux, or IBM i workloads.
Here’s what the Power7-based CloudBurst V2.1 setups look like:
IBM’s Power7-Based CloudBurst stacks. (Click graphic to enlarge.)
Assuming that you can do 10 partitions per core, and you have two BladeCenter chassis with all the storage and networking gear per chassis, the PS701 blades would have been a better choice for two reasons. First, there’s density, and then there is price. If you have 27 compute blades in the rack, just like you have in the X64-based HS22V CloudBurst setups above, you can cram a theoretical 2,160 PowerVM logical partitions in the same rack. You could argue and say that the PowerVM logical partitions are much skinnier than the VMware ESX Server 4.1 partitions, with only 1.6 GB allocated to the PS701 blades if you max them out to 128 GB. But if you match the memories on the VMs, you are still talking 64 LPARs per PS701 blade versus 36 VMs on the HS22V blades. IBM is allocating 256 GB of memory for 280 partitions on the Power 750s, which seems a bit skinny. It is very likely to get usable partitions, you will need to boost memory to 512 GB on the system, which gives to 1.8 GB per partition. But you need three racks, not one, to match what IBM could have done with the PS701 blades.
Like I said, it was an odd choice. Maybe it has to do with the Virtual I/O Server? Who knows.
The PS701 is a relatively inexpensive server, too, compared to the Power 750. The base PS701 blade with eight 3 GHz cores and 32 GB of memory costs $9,788 with no operating system. It costs another $11,394 to max out the memory on the blade and all of the cores come pre-activated on the blade, for a total of $21,182. The base Power 750 chassis in an Express configuration with all 32 cores activated and 128 GB of memory costs $101,952. Taking out the skinny 4 GB sticks and adding in the fat 16 GB sticks I think you need to push up to 512 GB of total memory raises the price up to $187,152 for 32 cores. Four PS701 blades with the same 32 cores, by comparison, would cost $84,728.
IBM made a bad choice in terms of density and price/performance. The Power 720 entry rack server would be a better choice to compete with X64 machinery. So if you are thinking of building AIX, Linux, or IBM i clouds, have IBM make you a custom cloud based on the blades or entry racks. The blades will give you the best density, and the PS701 and Power 720 will give you a better price than the Power 750 IBM is pushing. The Power 710 is disqualified from considering in my book because the eight-core versions are based on the expensive 3.55 GHz Power7 chips; the 3 GHz cores are not available in eight-core variants in this machine.
Anyway, the Power 750 variants of the CloudBurst stacks will be available on December 13. Pricing for the Power versions was not available as we were going to press.