April 4, 2011, Storage Network World, Santa Clara, CA—Dennis Martin from Demartektalk about the state of solid-state storage. In general, capacities are up and costs are dropping, and solid-state storage capabilities high enough to even encroach into the enterprise spaces.
An enterprise class HDD needs to perform between 100,000 and 200,000 IOPS while less than 100,000 is sufficient for a desktop drive. Depending upon controller architecture, the solid-state storage runs between 10 and 250,000 IOPS. Most solid-state storage is based on NAND flash and solid-state drives are available in capacities up to 1 TB.
The embedded technologies for storing data in flash devices use either single level cells or multilevel cells, and in both types writing is slower than reading. Single level cells are faster, have higher endurance, lower error probabilities, but cost more than multilevel cells. Most of the SLC devices are specified at 100,000 writes and unlimited reads, while a two level MLC is good for one tenth of that number.
The SLC drives cost about twice the cost of a MLC but the prices are dropping as manufacturers move to smaller processes with resulting smaller die sizes. The challenges are to get the error rate and reliability of even a SLC device up to those higher numbers as the processes go to 28 nm and below. Consumers and desktop computer users don't care as much about reliability as about cost, so a more comprehensive error correction capability is sufficient for a MLC drive.
SSDs use much less power than an equivalent HDD. An enterprise-class 15k HDD idles at 8-14 W and operates at 13-19 W. In comparison, an SSD may idle at less than 0.05 W and typically operates at about 3 W, with the full range of devices between 1 and 8 W. the applications for SSDs are in primary storage and as a cache for the HDD array. One trend is to automatically moving data so the "hot" data is available on the SSDs.
New designs for controllers are adding speed and enabling a single controller to handle multiple SSDs. The operating system has to be optimized for SSDs in order to enable background deletes such that the writes can take place immediately, instead of a block erase then write. New commands enable these behaviors, with "UNMAP" in SCSI and "TRIM" in SAS. Windows 7 can turn off background defragmentation to eliminate extraneous writes while Red Hat 6 only works on ext4 and the “trim” function must be enabled. The drive manufacturers also supply utilities to perform these functions.
A SSD in a cache configuration reduces the I/O loading on the HDDs by identifying the hot data and filling up the SSD with regularly accessed data. The SSD performance improves as the cache fills up, due to fewer cache misses. As a primary store, the user can identify the data and select what, when, and where data is stored. This results in an immediate improvement in performance for the selected applications, but needs constant monitoring to maintain that level of performance. Automatic software can help to select and move data to the SSD.
To demonstrate the differences in performance between HDD's and SSD's, they set up two different tests. The first was a Web server with a number of different storage configurations, and the second was as primary storage in a web environment. The requirements for the Web server are consistent response times, ability to handle large changes in traffic, and be cost effective.
The basic configuration for this test was a computer with Windows Server 2008 R2 with IIS 7.5. Default data compression was disabled and 8dot3 short names were removed and disabled. The test suite had 40GB of web server content comprised of 1.48 million files, 80,000 HTML text pages, and 1.4 million graphic images (JPEG and PNG). Altogether, this configuration represents a web hosting server with many sites and many pages.
Test 1 Caching O.S. and Web content
This test uses two different configurations. The first used 6 disk drives with the following specifications: 500GB SATA, 7200 RPM, 3.5-inch, RAID10. The second added 2 SSDs as cache: 32GB, SLC, 2.5-inch, SATA interface. In both cases the server connected to a 1 Gb ethernet network through teamed NICs. In a 90 minute test, the first configuration handled 1,285,708 total hits. Adding one SSD increased this number 3.3 times to 4,241,171 hits. Adding the second SSD increased total hits to 6,710,898.
The HDD configuration handled 58 Mb per second, the one SSD configuration achieved 211 Mb per second, and the dual SSD configuration hit 416 Mb per second. The throughput numbers flattened out for the first two configurations while the 2 SSD configuration was still increasing at 90 minutes. This implies that the SSD's were indeed offering the HDD outputs and as they filled up, more content was immediately available.
The most important consideration is average page response time. People perceive response times of less than one second as being inconsequential. Between one and 10 seconds, people consider the system to be slow, and over 10 seconds makes people drift away from the task at hand.
Adding an SSD to the storage array improves average response times
Test 2 Primary Storage Web content only
This test used three different configurations, the first one was 6 disk drives: 73GB 6Gbps SAS, 15K RPM, 2.5-inch, RAID10. The second one added more drives: 24 disk drives: 73GB 6Gbps SAS, 15K RPM, 2.5-inch, RAID 10. In the final configuration, the HDD's were replaced by 1 PCIe SSD: 300GB SLC Flash. The server was connected to a 10 Gb ethernet network.
In this primary storage test, the six drive system achieved 5,707,296 total hits in a 90 minute time frame. The HDD 24 configuration achieved 32,420,139 hits in the same time. The PCIe storage achieved 104,816,357 total hits. This demonstrates that SSDs and 10GbE go well together.
In the throughput portion of the test, the HDD-6 configuration averaged 310 Mb per second, the HDD-24 averaged 1681 Mb per second, and the PCIe card average 5429 Mb per second. The SSD was over three times faster than the HDD-24, and 17 times faster than the HDD-6. For the average page response times, HDD-6 was 0.177 seconds, HDD-24 was 0.033 seconds, and the SSD was 0.011 seconds.
If the only benefit to moving to solid-state storage was improved performance, these devices would be interesting but not compelling. One other area of comparison is power consumption.
Power consumption was less than 60 W for the PCIe storage, compared to over 100 W for the HDD-6, and over 200 W for the HDD-24
One additional consideration when contemplating the addition of solid-state storage to a system is that the bottlenecks will move to unexpected places. In the caching test, they found that a 1 Gb connection to the network was not fast enough. In all cases, CPU utilization increases, which could be a concern if the system is approaching capacity.
In the future, multilevel cell storage will go into the enterprise. SSD will become to your one storage by 2012. Looking out 3 to 5 years, be aware of research going on in other types of solid-state memory. Some of the other memory technologies include phase change, solid electrolyte, MRAM (magnetic), FERAM (ferro-electric), and RRAM (resistive or memsistor).