Getting Hyperconverged On Power Systems
July 17, 2017 Timothy Prickett Morgan
A few weeks ago, when pondering the possibilities of how Big Blue might reimagine and reinvigorate the IBM i platform, we outlined how this might be done – and how it might apply to a much broader customer base after it was done – in an article called The Cognitive Systems/500 2018 Edition. We forgot one of the possible components of a modern, integrated system and, amusingly, IBM has just announced support for it.
The bit of the software stack we forgot to mention was what is called hyperconverged infrastructure, or HCI for short, and this server-storage halfblood has become the underpinning of virtualized infrastructure at private datacenters for a number of key workloads and by thousands of organizations worldwide.
The idea with these HCI setups is fairly simple. Rather than have clusters of servers that support virtual machine hypervisors connect to usually expensive centralized storage area networks (SAN), HCI puts a distributed and virtual SAN on top of or inside of the hypervisor and the server nodes in a cluster do both the virtual compute and the virtual storage at the same time. These HCI stacks emulate, in a way, the sophisticated storage that was created by Google, Amazon, Microsoft, and Facebook a decade ago to underpin their consumer facing and cloud computing applications. The two most popular HCI stacks out there these days come from Nutanix, which has generated the most publicity and venture capital for this idea and which rolled out its first products way back in 2009, and VMware, the server virtualization juggernaut now owned by Dell that came to the market late but is growing fast because of the vast 500,000-strong installed base of its ESXi hypervisor and vSphere management stack at enterprises around the world.
There are other players in the HCI space, but these two represent the lion’s share of revenues and customers at this point. The Nutanix server-storage hybrid now runs atop its own implementation of the KVM hypervisor, called Acropolis, and it is a hodge-podge of different open source and proprietary technologies that create a very sophisticated distributed compute platform that is integrated and, at least compared to most open source stacks, easy to use. In those years, Nutanix has evolved from a strategy of “banning the SAN” from the datacenter to creating a true platform, and you can read all about the details about the success of the Nutanix Enterprise Computing Platform here and get a sense of the architecture there in two write-ups I have done at The Next Platform. The VMware stack is called Virtual SAN, or vSAN for short, and I have covered it in detail over here.
Both Nutanix and VMware have tuned their respective HCI stacks for the X86 server architecture, but there is no reason why either or both could not be ported to the Power chip. In fact, IBM has excellent support for KVM and has contributed its OPAL firmware and Power optimizations back to the KVM community, and this has enabled Nutanix and IBM to come to market with a Power version of the Enterprise Computing Platform. VMware has monkeyed around with supporting ESXi on ARM processors but has not, to our knowledge, ported it to Power. This is important because vSAN lives inside the ESXi hypervisor, unlike the distributed SAN that Nutanix has created, which runs atop its own or other KVMs, VMware’s ESXi, or Microsoft’s Hyper-V hypervisors inside of virtual machines.
IBM has created two models of the Nutanix server-storage hybrids, and it is interesting that these machines are the first ones to bear the Converged Systems banner from Big Blue. The CS821, formerly known as the Power S821, is a two-socket Power8 machine with a total of 20 cores and up to 256 GB of main memory that has four drive bays in a 1U chassis. The CS822 comes in a 2U chassis and has two Power8 chips with a total of 22 cores activated (that 11 cores per socket is weird, but that is what the specs say) and up to 512 GB of main memory and eight drive bays. Like other HCI solutions, the Nutanix stack uses a mix of flash and disk to boost the performance of the storage. The whole shebang runs Linux, and has the Acropolis hypervisor on top. The idea here is that is looks and smells like an X86-based server-storage hybrid, but it is running on Power iron, which has a lot more memory and I/O bandwidth and that should help the performance of virtualized applications running atop the hypervisor.
The pricing on these CS821 and CS822 systems will make your eyes water, but that is because somewhere around 90 percent of the cost of the system is the Nutanix software. The pricing on the X86 servers that Nutanix sells itself or through Dell, which was a Nutanix reseller before it bought EMC and VMware last year, are just as eye popping. The idea is that this software-based SAN mashed up with a virtual server layer is supposed to cost less than server clusters plus a real SAN, but to carry a nice, tidy profit while providing the very substantial operational benefits that come through integration of components and a platform approach.
Does any of this sound familiar?
The pricing on the CS821 ranges from a low of $43,500 for a system with 128 GB of main memory and four 480 GB SSDs to a high of $69,500 for a machine with 256 GB of memory and four 1.92 TB flash drives. On the more capacious CS822, pricing ranges from a low of $69,000 for a system with 256 GB of memory and eight 480 GB flash drives to a high of $115,500 for a machine with 512 GB of memory and eight 1.92 TB of flash. Note that IBM is selling flash-only configurations running Nutanix, but there are implementations available that mix disk and flash on X86 platforms from Nutanix and its partners.
The two Converged Systems platforms running the Nutanix stack will be available on September 1, and as far as I know, IBM is not allowing for Nutanix to run on logical partitions where AIX and IBM i workloads are sitting side-by-side. But you can guess what I am about to say here.
First of all, if Nutanix can be a virtual SAN for KVM instances running on Linux on Power, there is no reason that it cannot be a kind of hyperconverged Integrated File System for IBM i, much as the actual Integrated File System from OS/400 V3R6 was based on OS/2 Parallel File System that was quite good back in the mid-1990s when IBM borrowed it to give ASCII files and desktop server operating systems something familiar for them to talk to. I see no reason why IBM i and AIX cannot be propped up on the OPAL firmware and the OpenKVM hypervisor, and therefore also be able to support Linux partitions on a single server or a cluster of servers that could support IBM i and AIX in virtual standalone mode above a distributed virtual SAN on a bunch of Power-based systems. What is good for Linux is good for IBM i and AIX, and in fact, porting these two platforms from the earlier PowerVM hypervisor to OpenKVM and its Acropolis KVM brother makes a lot of sense given the fact that HCI platforms are aimed specifically at midrange shops that do not have a lot of IT people.
If the high availability software vendors who dominate the IBM i market wanted to branch out into a new area, this would be a good one because a distributed, clustered server-SAN hybrid has many of the attributes of a HA pair or triplet running some of their HA software. These companies have the muscle to get IBM to do the porting work, and this would give them a chance at selling a solution that is consistent across IBM i, AIX, and Linux and not different for each platform. There is some profit margin in here, too. That said, going down this road would mean abandoning existing approaches and products, and that is probably a scary thought. But it would be nice to have a solution that can scale across many nodes both on site and across sites, and this is not a simple task today. Products like Nutanix Enterprise Compute Platform help you make a baby Google in your own dataclosets and datacenters.
It’s something to think about, and it can be woven into the proposed stack I made for the theoretical Converged Systems/500 machine I was talking about last month.