IBM Melds FlashSystem With XIV To Scale Enterprise Flash
May 9, 2016 Timothy Prickett Morgan
Over the past few weeks, we have been analyzing the new flash modules and controllers that IBM has made available for plugging into Power Systems machines and significantly boosting the performance of the storage arrays under the skins of the servers. This week, we will take a look at the new FlashSystem all-flash arrays from Big Blue, which are aimed at much larger workloads that cannot fit within the system.
The two new FlashSystems that IBM announced concurrent with the internal flash storage back in April are much heftier machines than it has been peddling to date, and the more interesting of the two is actually a mashup if its FlashSystem all flash arrays and its XIV clustered disk arrays, which are now referred to as Spectrum Accelerate. IBM, you will remember, acquired XIV back in January 2008 for somewhere between $300 million and $350 million. Back in 2011, when the XIV line was upgraded, IBM had over 1,100 customers with more than 4,500 unique XIV clusters installed, and as of this year, the XIV software, which can run on IBM’s appliances or on X86 hardware supplied by customers, is installed at 6,000 customers and runs on an installed base of more than 100,000 servers. That is a lot of capacity, and some of it is driven by the IBM i operating system, which has been supported by the XIV clustered storage since July 2009.
The FlashSystem is also supported by IBM i, and a compelling case can be made for the use of FlashSystem storage plus the SAN Virtualization Controller (SVC) mashed up to provide virtualized storage with snapshotting and data compression. (The FlashSystem did not provide these functions, so it was added on with the SVC, which also added a bit to the latency of data reads and writes, as any controller running such software does.)
The FlashSystem A9000 is a follow-on to the existing FlashSystem V9000 that IBM has been shipping for some time, and it crams three different controllers that run the Spectrum Accelerate software and a dozen FlashCore flash modules that come in 1.2 TB, 2.9 TB, and 5.9 TB capacities; these are based on 20 nanometer enterprise MLC flash chips from Micron Technology. The controllers each have two eight-core Xeon E5 v3 processors running at 2.4 GHz running their software (it would be nice if this was based on Power chips) and 192 GB of DDR4 main memory. The controllers are set up in a three-way active grid to spread out the Spectrum Accelerate software and make it run better and survive failures. The FlashSystem A9000 can serve up to 500,000 I/O operations per second with a 250 microsecond latency and delivers a maximum throughput of 4.5 GB/sec. The raw capacity of the FlashSystem A9000 ranges from 21.4 TB to 105.6 TB, and assuming reductions from data compression, thin provisioning, data compression and the overhead of RAID 5 data protection across the flash modules, the effective capacity cranks up by a factor of 5.26 to 1 on average for a range of 60 TB to 300 TB. By the way, the nodes are linked with 56 Gb/sec InfiniBand links, and data encryption using 256-bit AES-XTS algorithms is built in. Hosts link to the A9000 through 16 Gb/sec Fibre Channel or 10 Gb/sec or 40 Gb/sec Ethernet links using the iSCSI protocol.
By the standards of the latest all flash arrays, this is neither compact nor capacious nor particularly fast. EMC Nitro arrays, which will ship next year, and Pure Storage FlashBlades, which are expected by the end of this year, are denser and faster. The FlashSystem A9000 is important for IBM i shops in two important respects, though: it is available now, and it supports IBM i.
If you want scale, then the FlashSystem A9000R is the ticket. With this setup, the XIV clustering software is turned on so that two FlashSystem controllers and a flash drive enclosure can be clustered to scale out further than the single A9000 node can do. In this case, up to six enclosures sporting 2.9 TB and 5.7 TB drives can be lashed together into a single FlashSystem image with a maximum raw capacity of 316.8 TB and 633.8 TB, respectively, and an effective capacity of 900 TB and 1.8 PB, respectively. The A9000R with six enclosures can drive up to 2 million IOPS and has a maximum bandwidth of 18 GB/sec while still maintaining that 250 microsecond average latency. The nodes in the cluster are linked via 56 Gb/sec InfiniBand as well.
You can drill down into the feeds and speeds of the FlashSystem A9000 and A9000R in this Redbook that IBM just published. What IBM does not make very clear is that the new FlashSystems can back-end various IBM i, Linux, Unix, and Windows servers and can also support the popular hypervisors on these machines, namely PowerVM and PowerKVM on Power machines and KVM, Hyper-V, and ESXi on X86 machines. The capacity on the FlashSystem can be carved up to give specific slices to particular operating system instances, which keeps everything from getting scrambled. The FlashSystem A9000R might be particularly interesting for service providers wanting to build out high performance clouds that include IBM i and AIX instances running on Power Systems as well as Windows and Linux instances running on X86 iron.
IBM is providing pricing for the A9000 here and for the A9000R there. IBM says that pricing is as low as $1.50 per effective GB of capacity, which is in the same ballpark as other all flash arrays these days, and we presume is subject to variation based on configuration details.