ION recently launched it’s flagship SR-71mach5 SpeedServer™ which continues to deliver on the SR-71 legacy of all-SSD storage for “disk-impossible performance”. This new server incorporates new technology throughout yielding more processor cores, more and faster memory and much more random I/O performance.
Standard processors in the “mach5” are two 12-core 2.6GHz Intel® Xeon® E5-2690 V3, but the full Xeon E5-2600V3 processor lineup is available. The serverboard has, of course, been updated to support the V3 processors and includes (2) 10GBASE-T ports and more PCI Express slots. The system now supports up to 512GB of 2133MHz DDR4 memory or up to 768GB at 1600MHz. The new RAID controllers are Intel® RS3 series supporting 12Gb SAS3. Finally, the featured SSDs are Intel® DC S3610 series with Medium Endurance Technology rated for 3 drive writes per day for five years. While the SSDs have a 6Gb SATA interface, the SAS3 RAID controllers support many more IOPS and higher throughput than their SAS2 predecessors. Capacity of a single 2U SR-71mach5 SpeedServer is now over 25TB!
New updates to the SR-71mach5 SpeedServer Configurator now offer choices for both Standard Endurance (S3510) and High Endurance (S3610) Intel SSDs, both internally and in expansion shelves. Options for Intel® PCI Express NVME SSDs have been added, as have options for Intel® Xeon® Phi™ CoProcessors.
The “mach5” server maintains the disk-impossible random I/O performance of earlier generations, while posting dramatic improvements in average latency. This article details some of those improvements for “on target” I/O.
Random read operations of 4kB blocks on the SR-71mach5 SpeedServer, as measured by IOmeter on Windows Server 2012 are performed at about 1.4M IOPS (Input/output Operations Per Second) but the average latency measured during those results decreased from about 330 microseconds (mach4) to 240 microseconds(mach5), or 27% lower. Tweet
Performing 64kB random write tests with IOmeter, the new “mach5” server wrote at over 21,000 IOPS or 1.35MBps with 9ms latency when queueing 8 outstanding I/O requests. With just 4 requests queued, the system still performs over 16k IOPS or 1MBps with an average latency of just 5.7ms. IOPS have increased and latency has decreased compared to the previous generation.
The IOmeter OLTP profile performs random read and write operations at a given block size, with a mix of 1/3 write with 2/3 read. Running the 64kB OLTP test, the new “mach5” server with a queue depth of 8 performed 42k IOPS with average latency of 4.6ms while the mach4 server with a queue depth of 20 performed 18k IOPS with an average latency of 11.3ms. Keeping performance constant with fewer outstanding I/Os queued reduced the average latency by almost 60%.
OLTP testing with 8kB blocks showed even greater improvements in the mach5 system compared to the mach4 server. With a queue depth of 14/20 IOPS increased modestly from 243k to 247k, but average latency dropped 30% to 1.35ms. With a queue depth of 8, the server still delivered over 200k IOPS with a sub-millisecond latency of just 940 microseconds.
Microsoft’s SQLIO is a tool for evaluating hardware for usage with Microsoft SQL Server. Improvements in both number of IOPS and average latency were noted with SQLIO as well. 8kB random writes are now performed at over 100,000 IOPS with average latency of just 1ms. 64kB random writes occur at over 18k IOPS with just 11ms average latency while 64kB random reads are up to 144k IOPS with average latency under 1ms.
This article reiterated ION’s description of SR-71mach5 SpeedServer performance as “disk impossible.” To achieve 1.4 million 4kB random reads in a system based on 10k RPM 2.5″ disks capable of about 300 IOPS each would require more than 4666 disks. That would be 194 RAID subsystems occupying 2U each (that is (10) 42U racks) and drawing 350 Watts each, or 68,000 Watts total. The SR-71mach5 SpeedServer occupies just (2) rack units and consumes less than 500W. And then compare latency of 0.24ms with idealistic single-disk latency of 2.9ms. Even if you pretend that latency does not increase when you scale out to 4600 disks, average latency on the SR-71mach54 server is 91% lower. If you compare based on a read/write test like 8kB OLTP where the “mach5” performs about a quarter of a million IOPS, you would still need at least 800 disks or (34) 2U RAID subsystems, assuming RAID 5 or 6. So, whether you have a limit on acceptable latency, or rack space, or power consumption, or cost, it is impossible to reach these performance levels with a solution based on spinning disks.