AMS 2000 – Not Your Mother’s Modular Storage – SPC-1
by Hu Yoshida on Mar 24, 2009
Modular storage systems were developed over twenty five years ago. They were developed as external storage for UNIX computer systems that used a simple SCSI, small computer storage interface for I/O processing. The disks were formatted in a Fixed Block Architecture. This was a departure from Mainframe storage systems of that day that used channel processors that processed I/O using channel commands that were executed on disks that were formatted with a Count Key Data format.
While mainframe I/O was designed from the early beginning to handle a virtual server operating system, clustering, shared I/O, multiple formats, and process I/O on the fly, modular storage was designed to be very simple to configure and implement in a direct attach environment at a much lower cost. Initially they only had one controller. Later a second controller was added for availability in the event of a failure in the first controller.
Since each controller had its own cache they could not work on the same LUN at the same time without cache coherency problems. A LUN had to be assigned or owned by an “active” controller while the other controller was kept in a “passive” mode, ready to take over the LUN in the event of a failure in the first controller. Some vendors would claim that their dual controller modular storage was active/active. This meant that each controller was active, but on different sets of LUNs. While there were a few major enhancements over the years like RAID and FC, the basic architecture of modular storage systems remained the same until recently.
In 2008, Hitachi Data Systems introduced the AMS 2000 modular storage family with the first dual active controller architecture. This modular storage system no longer requires LUN ownership by one controller or the other. I/O can come down either controller path to access a LUN in cache without any consistency problems. If one controller fails or is taken down for maintenance or upgrades, the other controller takes over the workload. If one controller starts to become busier than the other controller, the workload is automatically rebalanced. This makes it much easier to configure, maintain, and tune for performance and availability.
In addition to the high performance dual active controllers, the AMS 2000 has full duplex, 3 Gbs SAS drive interfaces which can simultaneously send and receive commands and data on the same link. With SAS expanders it can support 32 concurrent disk I/O paths. This compares to other dual controller modular storage systems that only have FC arbitrated loops that are shared across all the disks.
With these two major enhancements, the AMS 2000 series of storage systems has changed the playing field for modular storage. One measure of these enhancements is revealed in the SPC-1 results which were released today:
In the SPC-1 benchmark testing the AMS 2500 achieved the fastest throughput results among all modular dual controller storage systems with a result of 89, 491.81 SPC-1 IOPS and an 8.98 ms average response time.
In the SPC-1 Price-Performance category, the AMS 2100 achieved among the best ratings in its class with $5.95 per SPC-1 IOPS
For more information on the SPC-1 Results for the Hitachi AMS 2000 family, visit:
AMS 2500: http://www.storageperformance.org/results/benchmark_results_spc1#a00078
AMS 2300: http://www.storageperformance.org/results/benchmark_results_spc1#a00077
AMS 2100: http://www.storageperformance.org/results/benchmark_results_spc1#a00076
Comments (5 )
Good data, Hu.
The alternative to the mainframe of twenty five years ago was the minicomputer, most of which did not run Unix. The departure from the mainframe storage model goes back much further to the founding of the minicomputer industry c. 1960 from which Unix was later born.
Mainframe storage (parallel storage channels and processors) was originally designed simply to offload the device management from the central processor, which was too expensive for such work. Also, CKD was a System/360 era invention of c. 1965, no one had yet thought of virtualization or clustering.
Minicomputers, with their much cheaper central processors, could get by managing their own IO and saving the price of specialized PIO devices. A nice synergy that survived from the beginning until the 1980s when outboard disk channels and microprocessor based disk controller boards emerged for minicomputers.
The development of the micro based controller boards led to the adoption of SCSI, which brought the invention of today’s storage arrays.
[...] Yoshida recently blogged on the latest AMS SPC-1 results and the numbers are as you can expect -impressive. There is often heated debate over the [...]
Charlie, thanks for the information. Do you have more that you can share with us about the history of storage? Are there any reference materials that you can point us to? While there is a lot about recording technologies, there doesn’t seem to be much written about storage systems.
I agree, a history of storage is needed while the inventing generation is still with us. I joined DEC out of school in 1983 so what I know from then on is experience. Prior there are some good books, though company specific and not entirely about storage.
The information about the cleavage between minicomputers using DMA and interrupts vs. mainframes using channels and outboard controllers is documented in a DEC book, long out of print, called “Computer Engineering: A DEC View of Hardware Design.” Gordon Bell, one of the authors, has preserved it on line at Microsoft where he now works. The specific reference can be found at
I went back and re-read it to write this message. The idea is somewhat older than I had thought! The CPU based IO handling goes back to the MIT TX-2 in c. 1957 which influenced the DEC PDP-1.