Author: Kent Bransford, Senior Technical Editor,
The following is an actual case study, illustrating how SAS technology is a key enabler of high-performance business intelligence storage solutions that utilize the concept of a “server brick.”
Microsoft Project REAL
The server brick concept was developed by Microsoft while working on Project REAL (Reference implementation, End-to-end, At scale, and Lots of users). The project involved a major US retailer with over 800 stores in the US and more than 40,000 employees. While the retailer’s sales and data warehouse was initially 1.5TB in size, by the project’s completion it had grown to over 4TB and was expected to continue this rapid expansion in the future.
The retailer’s application usage model was characterized by an enormous number of database queries, of which 95% involved accessing sequential data. The original system design was a conventional SAN solution, which proved difficult to scale in order to meet the retailer’s performance objectives. Due to the sequential nature of the application, the system design requires a tremendous amount of bandwidth between the servers and storage.
Project REAL’s goal was to develop a system that would deliver superior raw bandwidth to the server in a very small footprint.
Figure 1: Retailer’s Existing System Footprint (left) vs. Server Brick Footprint (right)
Project Hardware: Prototype Server Brick
The initial system configuration consisted of 24TB of SATA storage (64 400GB disk drives) connected to a 4-socket dual-core AMD processor with 32 GB of memory. Six SAS controllers and four expanders provided connectivity to the storage. All components were housed in a 15U rack-mounted design which satisfied the small footprint criterion. Total cost of the disk storage and supporting hardware came to less than $2.40 per GB (non-discounted pricing).
Figure 2: Project REAL’s Reference Prototype Server Brick
Additional cost savings came from reduced power consumption, with the SAS solution using approximately one-tenth of the power needed by the large SAN it replaced. This reduction in power requirements is partly attributable to the SAS solution’s use of less memory and fewer processors. Complete cost of the SAS system was less than $50,000, far lower than the existing SAN solution that failed to meet the retailer’s requirements.
Test results: Prototype Database Server Brick
The application was running an SQL Server which provided the benchmark test routines to measure the effectiveness of the SAS server brick. As can be seen from the test results below, the server brick outperformed the SAN solution in all test cases save one. Large-range query performance differences were primarily due to the large amount of DRAM utilized in the SAN solution, and not the direct result of drive performance or storage connections.
The main factor influencing the other benchmark results was the significantly higher bandwidth enabled by the SAS wide links to the server.
Figure 3: BI (Server Brick) vs. SAN Test Results
Benefits of Database Server Bricks
Rich Johnson, Business Intelligence Architect for Microsoft Consulting Services, considers server bricks a major paradigm shift in performance, capacity and price models compared to traditional solutions. For example, server bricks maximize the value of the relatively few PCI-X or PCI-e slots available in small 4-way servers. SAS architecture enables them to deliver extremely high throughput to the limited number of PCI slots in the servers.
Even when utilizing small servers and handling terabytes of storage, the server brick solution proved it can deliver higher performance than large SMP SAN solutions costing ten times as much.
Compared to traditional datacenter SAN solutions, direct-attach SAS storage is now a viable, economical alternative for many data center applications, both today and in the future. Combining superior random I/O performance with a space-saving form factor, direct-attach SAS solutions open up many new opportunities in data warehousing, including applications such as streaming video, eCommerce, data mining and OLTP (On-Line Transaction Processing).
Today’s storage landscape contains an enormous mid-market that is woefully under-served simply because many such customers can’t afford multi-million dollar SANs. With the advent of affordable SAS-based server bricks, companies can now get SAN-class performance—or better—for a fraction of the cost.
Following the initial tests noted above, the SCSI Trade Association designed an updated system and performed additional benchmarks. The new system included servers with dual-core processors and PCI-e slots, each equipped with two SAS RAID controllers. These were connected to a SAS switch that in turn attached to five JBODs from various suppliers, housing a mixture of 2.5-inch and 3.5-inch disk drives (both 10K &15K RPM).
Note that because SAS is an industry standard, there are numerous suppliers that manufacture interoperable components that can meet any application requirements.
Tests using SQLIO throughput benchmarks produced compelling results:
- 3.2Gbps (3,200MB/s) for 64KB sequential reads
- 2.7Gbps (2,700MB/s) for 64KB sequential writes
- 2.0Gbps (2,000MB/s) for 256KB random reads
- 2.1Gbps (2,100MB/s) for 256KB random writes
The net result? Superior performance in a very small footprint, all at a price point amenable to many different application needs and budgetary constraints.
Additional tests were run using sequential reads and writes of short block transfers (512 byte blocks), in order to see how many I/Os per seconds the system could deliver. The system was capable of executing over 500,000 I/Os per second, a remarkable performance that not only proves SAS technology can deliver extremely high throughput, but also efficiently serve applications that entail intensive I/Os per second such as OLTP.