In an article titled "No Waiting: Considering the Benefits of Solid State Disks,"1 authors Ramon Sandoval and Maneesha Lee highlighted the growing popularity of solid state disks (SSDs) as accelerators for enterprise databases. The authors cited the rapid expansion of the SSD industry as a result of relational database applications, with SSDs being deployed as storage devices for hot files2. In particular, Sandoval and Lee revealed that a major portion of enterprise SSDs are installed in high-end database applications running Sybase, Oracle and Informix, along with SAP running on top of these databases.
Aside from this write-up, several reports and white papers emphasizing the performance benefits of SSDs have already been published, and hard disk drive makers such as Samsung have expressed interest in this storage line. The only remaining question is, are SSDs a viable alternative for price-conscious buyers who are running performance-hungry OLTP apps?
The objective of this article is to examine the benefits of utilizing flash SSD-enabled storage system as cache storage in an enterprise environment. A comparison will be made between conventional storage systems (featuring HDDs) and solid-state disk-based network storage in terms of performance and overall cost per IOPS.
Market Trends
Storage users are buying more midrange and lower cost systems, according to International Data Corp.'s (IDC) quarterly worldwide disk storage systems report released December 2004. Revenues grew 3.5 percent year-over-year to $3.4 billion in the third quarter of 2004. Although revenue growth was smaller compared to previous quarters, IDC noted rapid growth in storage capacity, rising 50.5 percent year over year to 310 petabytes. It is the largest growth rate posted over the last seven quarters, the report said.
IDC analyst Brad Nisbet says the results confirm the slow but steady growth of midrange and lower segments, like ATA-based storage. "We saw an increase in the growth of petabytes shipped, which is yielding the largest dollar per gigabyte pricing decline in seven quarters and points to a growing share of higher-capacity, lower-cost disk drive deployments and a broader variety of products offered by the major vendors," he explains.
Another contributing factor to the growth of networked storage, in particular midrange systems, is the stiff price competition among sellers in their bid to corner a slice of shrinking IT budgets. Price consciousness is also echoed in the way enterprises handle their storage requirements. Consolidation is now a must as it increases storage manageability, maximizes capacity utilization, and lowers overall TCO.
However, this strategy puts a strain on server performance, especially for I/O intensive applications such as:
Aerospace, telemetry and data acquisition
Data backup and recovery as well as migration
Energy exploration and geosciences
Medical sciences, including healthcare and imaging
Online transaction processing (OLTP)
Paging, log, journal and index files
Still and moving video surveillance
Video editing and processing, including post-production
Video on demand and video services
Weather forecasting and simulation
OLTP is a critical segment in enterprise storage as most banking, trading and supply chain transactions are now transferred online, with users demanding faster and more responsive systems. In selecting the most cost-effective, high-performance storage device for this application, IT and data center administrators have HDDs and SSDs as options. The wide variety of available models in the market and lower price per unit makes the HDD as the most logical solution for enterprises, providing the best return on investment. Or does it?
Cost Analysis
Let us examine a setup involving a generic SAN attached JBOD setup comprised entirely of 15,000RPM rotational HDDs as opposed to a BiTMICRO E-Disk®SAN S2F-J featuring flash-based E-Disk SSDs. The objective of this analysis is to compare the capital investment needed to acquire a networked storage solution that can offer peak performance of 100,000 IOPS in a typical OLTP/database I/O application.
Hardware Specifications
The tables below detail the technical specifications of the drives and JBOD modules used in our comparative analysis.
In this study we have selected 15,000rpm drives, the fastest available rotational hard drives in the market today. Both drives feature Fibre Channel interfaces and are enclosed in the same 3.5-inch form factor. Since this analysis will focus more on cost versus IOPS performance alone, disk capacity is totally irrelevant in this analysis.
To cancel out the advantages (and disadvantages) brought by JBOD modules from different manufacturers, this analysis will utilize a generic JBOD module that has the same form factor (2U) and disk capacity (12 3.5" disks) as the E-Disk®SAN S2F-J from BiTMICRO.
Disk/Module Performance
There has been a dearth of storage hardware literature that tackles HDD performance as measured in I/Os per second, as most drive manufacturers publicize disk performance in terms of MB per second (MBps). However, IOPS statistics are critical in random access applications such as OLTP and data warehousing, and storage subsystem suppliers post mission-critical IOPS data in their website. Similarly, BiTMICRO Networks conducted benchmark tests in November 2003 for its E-Disk®SAN featuring twelve E-Disk Fibre Channel channels. IOPS results for a small-block (4KB), sustained random read workloads are posted in the succeeding table.
To compute for the generic JBOD's performance rating, we multiplied the 15,000RPM HDD's maximum IOPS rating of 435 with the maximum number of disks in the enclosure (12). The result, 5,220, is the theoretical maximum IOPS for small block (4KB) random reads.
Table 3 figures already show a wide disparity in I/O ratings (more than 1100%) both at the drive and at the enclosure levels. To achieve the desired performance of 100,000 IOPS, we simply add more enclosures and drives in the JBOD setup accordingly (Table 4).
To compute for the dollar cost per IOPS, we add up the equipment cost and divide the amount by the total IOPS generated. Equipment includes the drives, enclosures, switches and cabinet. However, for the purpose of this analysis, it is assumed that each setup will utilize one switch. Therefore the cost of switches may not be included in the computation since they would just cancel each other out. List prices for the generic JBOD setup are based on web prices as of June 29, 2005.
Analysis of Results
To compute for the $ cost per IOPS, the following formula will be used using figures from Tables 4 and 5.
Total IOPS / Total cost = $ per IOPS
Table 6 figures reveal a significant difference in investment costs when building a storage system for database/OLTP applications using a generic JBOD versus an E-Disk®SAN. On a $/IOPS level, the rotational HDD-based JBOD array costs 303.8 percent higher than the E-Disk®SAN, thanks to the significantly higher I/O rating of the S2F-J solid state disk subsystem which resulted in a lower number of drives required. Another factor that drove down equipment cost is the smaller cabinet space requirement of the S2F-J (24U) as opposed to the generic JBOD (48U). In fact, the SSD-based subsystem can be easily integrated into existing data centers since it is composed of only two 2U modules as compared to the 20 modules required for the HDD-JBOD setup. Removing the 24U cabinet (costing $4,298) in the S2F-J setup would drive down $/IOPS even further.
Conclusion
Enterprises are looking for ways to generate more IOPS in the same data center space, as well as cost-effective means of scaling their storage performance requirements. The examples in the preceding section prove that solid state flash disks are cost effective solutions for performance hungry applications, not only in terms of $/IOPS in general, but also for deferring costly server upgrades in I/O bound applications. This article focuses mainly on the cost advantages of the SSD subsystem in an equipment level. The gap between SSDs and HDDs widens further when other factors such as power consumption of 220 rotational hard drives versus 21 solid-state flash disks, storage space (and the burden that these subsystems will impose on the data center's cooling system), and better reliability (MTBF) are taken into consideration.
------
1 Source:
http://www.xchangemag.com/articles/0b1feat3.html2 Defined by techweb.com as "a database or data file that is currently being heavily accessed by users"
