|
|
Flash Solid State Disks: A New Breed of Enterprise-Class Storage
By Jun Alejo, BiTMICRO Networks
As seen from Figure 3, SSDs start to lose this advantage in larger block transfers, which is specifically lost in purely random writes of 128KB or larger blocks typical of data streaming applications. However, this workload profile is not characteristic of certain multi-user streaming applications, for example a video-on-demand (VoD) server.
Note: Figures 3 and 4 are based on BiTMICRO Performance Lab Tests dated 02/05/05 comparing the E-Disk Ultra320 solid-state flash disk vs. the Seagate ST336753LC Ultra320 80-pin, 15,000rpm HDD
In addition, note that the said disadvantage will be a thing of the past as the next wave of E-Disk SSD solutions are unleashed to the market, featuring 4x the current maximum sustained transfer rates as well as sustained writes that are finally at par with that of reads.
Most of the implementations will be on RAID-0, 0+1 or 1+0, as long as it is striped. IOmeter tests on a single device will not simulate such an environment because although the access is random, the I/O read request will still wait for the device to complete reading the whole block (typically 64KB or greater). This points to the fact that a very small amount of time will be spent on seeking, when compared to the amount of time to read the data. In this situation, we do not take advantage of the fact that HDDs need to seek every now and then.
On the other hand, a multitasking, multi-user environment fragments and buffers I/Os as the OS services different read requests from a number of users, sending all heads of the RAID matrix disks in different directions. The main difference is that the whole block of data requested by the user is not necessarily read in one pass, especially if it is a large video file for example. This situation is analogous to random read patterns of smaller (by a factor of the number of disks in the RAID set) block transfers in an IOmeter benchmark environment off a single disk, smallest block of which is the strip size of that disk. This is the reason why SSD performance rises above HDDs in environments where there is more than one user/host.
Having more than one IOmeter host as drivers to SSD devices better simulates a multi-user/multitasking environment, highlighting SSD's advantage over HDDs in terms of IOPS. Given the fact that most, if not all, storage arrays for data streaming applications will be striped, an IOmeter benchmark of purely random reads of smaller block sizes is therefore a more realistic representation of the I/O patterns of a RAID matrix disk on a multi-user video server storage application.
SSD Solutions for Networked Storage
Performance - In previous years, most enterprises resolved performance issues by increasing the number of storage servers or extra disks to boost CPU cycles or ramp up the number of I/Os per second. However, such method requires additional capital outlay to shoulder the cost of additional storage servers and disks. This solution also increases the complexity of the storage network, leading to a higher cost of monitoring and maintaining the system.
As mentioned in the previous section, certain applications stand to profit from SSD deployment. Business applications that exhibit random access patterns and small to medium block I/O (2KB - 64KB), such as online transaction processing (OLTP), e-mail, Unix file system (UFS), e-commerce, and common Internet file system (CIFS), would benefit most from the fast access times of SSDs. By transferring frequently accessed data (also known as "hot files") to an SSD such as the E-Disk® Flash SSD, the need for subsystem upgrades are minimized, resulting in a significant reduction in total cost of ownership. Overall RAID performance also improves since most I/O requirements are handled by a dedicated disk.
What are hot files? These are frequently accessed files such as passwords, system and application logs, journals, and temporary tables. According to IMEX Research, a mere 2 percent of corporate data represents 65 percent of all I/O access in an enterprise. The impact and cost benefit of installing an E-Disk Flash SSD as a hot disk1 becomes more significant as you target "regular" data normally resident on disk arrays towards the lower left side of Chart 5, an area representing "hot files."
Balanced I/O Computing - As illustrated in Figure 1 of this paper, storage transfer speeds have not been able to keep up with the exponential increases in CPU speeds. Server computing at present is concentrated mainly on I/O processing as opposed to CPU processing. In 1995, typical application workload distribution is divided equally: 50 percent of the time is spent in processing while the other half is spent accessing data.
However, in the year 2000, servers spend 90 percent of their time in I/O processing2. Therefore, investments in CPUs will only result in negligible improvements in system performance, as opposed to exponential improvements resulting from the installation of extremely fast Flash-based SSDs.
1 Storage Networking Industry Association defines "hot disk" as a storage device whose capacity to execute I/O requests is saturated by the aggregate I/O load directed to it from one or more applications
2 "Balancing Server CPU & I/O w/ FlashDisk RAID", Winchester Systems
[ Top ]
