WHITE PAPER Amadeus SAS Specialists Prove Fusion iomemory a Superior Analysis Accelerator 951 SanDisk Drive, Milpitas, CA 95035 www.sandisk.com
SAS 9 Preferred Implementation Partner tests a single Fusion iomemory iodrive device against SSDs and local storage and finds that the Fusion iomemory iodrive device greatly outperforms both. Executive Summary Continual growth in data volumes presents great challenges to the performance of SAS computing environments. Most often, the reading and writing of data to storage, referred to as I/O (Input/Output), is the biggest performance challenge. Amadeus Software, a SAS 9 Preferred Implementation Partner, designs SAS solutions to meet any need. They conducted tests that stored data on a Fusion iomemory iodrive device, an array of SSDs, and a hard disk to compare the media s effectiveness at reducing I/O wait time and improving end-application SAS performance. Amadeus Software observed an average reduction of 58 percent in the elapsed time of SAS jobs running on Microsoft Windows Server 2008 R2 (x64) servers when using the Fusion iomemory iodrive device as compared to hard disk, much greater than the 38 percent improvement for an array of SSDs. With the Fusion iomemory iodrive device, Amadeus observed individual procedure steps peaking at a five-fold performance gain when compared to the mechanical disk. Overview The raw volume of data stored by organizations is growing exponentially, increasing the workload of jobs. At the same time, SAS users are using expanded software capabilities that increase job complexity, all the while demanding faster results. The primary problem with traditional SAS systems isn t slow SAS application performance, but the inability of hard disks to cost-effectively deliver the I/O needed to keep SAS CPUs productively utilized. The emergence of NAND flash technologies offers orders-of-magnitude faster I/O performance, but there are differing approaches on the market. Amadeus Software ran tests that compared two solid-state technologies: Fusion iomemory s flash-based memory tier and an SSD flash array. The intent of this study was to determine which is most effective at improving application throughput. It discusses the following topics: About the Solid State Technologies About the Tests Test Results About the Solid State Technologies Fusion iomemory iodrive Devices Instead of approaching NAND flash like a hard disk, Fusion iomemory iodrive devices seek to maximize CPU utilization by minimizing latency (the distance between the storage and the application). Fusion iomemory Virtual Storage Layer (VSL) maps application block requests directly to NAND locations on the device, similar to the way data is accessed in memory page tables. The goal is to achieve the highest levels of application throughput by eliminating the overhead and bandwidth limitations of storage protocols and onboard processing associated with disk form-factor SSDs, as well as the deep queues that add latency and divert CPU resources to managing threads and contexts waiting on I/O. 2
Flash-based Storage Array The second solid-state technology will be referred to as Solid-state Technology B. It connects a flash-based storage array to the server over a SATA (serial ATA) port. This device s approach to performance is to imitate a SATA disk and then aggregate performance across multiple SSDs in an array. Hard Disk Drive Amadeus Software also ran tests against the standard mechanical hard disk found in most servers NAS or SAN arrays as a control group. About the Tests Amadeus s tests represented typical SAS programming tasks. Amadeus submitted each test program at least three times in an independent batch SAS session and averaged the results. Amadeus created a number of SAS programs that manipulated data sets with a variety of DATA and PROC steps. They executed each program on a range of data set sizes as a single SAS session, and as one of a number of concurrently executing SAS sessions data sets that consisted of tables of sizes: 0.5GB, 1GB, 5GB, and 10GB, containing variables, with an equal spilt between character and numeric. Character variables contained random sequences of 255 alphabetic characters, while numeric variables (8 bytes) contained random numbers between 0 and 100. Each SAS program contained a number of single-threaded and multi-threaded procedures. Multi-threaded procedures were Proc SORT, Proc SQL, Proc SUMMARY, and Proc TABULATE. SAS compression was not used. Programs with and without SAS indexes were used. The following is a brief description of each step in the program: The data step read from an existing data set. An indexing function searched each character variable for a string of characters. The results were then output to a new data set in the work library. Proc SORT read from the original table and sorted the table by a variable, outputting the result to a new table stored in the work library. Proc SQL read from the original table. The table was then ordered by a variable. The program then selected all of the variables and placed them in a new table written to the work library. Proc TABULATE produced a cross tabulation, classified by two class variables and calculating the mean of a numerical variable. The results were also written to an output data set in the work library. Proc SUMMARY read the data from the original table, ran summary statistics, and printed the resulting data set to the output window and to a new table written to the work library. Proc TRANSPOSE read the data in the original data set and created a new data set to the work library, transposing the rows to columns and vice versa. Finally, Proc CORR read from the original data set and calculated Pearson s correlation coefficients on all of the numerical columns, outputting the matrix to a new data set in the work library. 3
Test System Test server: 8 CPU cores, 32GB RAM OS: Windows 2008 R2 (x64) Application: SAS 9.2 (TS1M0) Test Results Table 1 below summarizes the performance benefits of the Fusion iomemory iodrive device, the SSD array, and a traditional mechanical disk. These results are from tests with a 10GB input data set consisting of 150 character variables and 150 x 8 byte numeric variables, with tables uncompressed. Overall, the Fusion iomemory iodrive device was found to be 58 percent more efficient than the mechanical device for a 10GB data set, compared to 38 percent faster for the aggregated SSD array (Solid State Technology B). Real Time (seconds) Performance Increase (%) vs. Mechanical disks Fusion iomemory SSD B Mechanical Fusion iomemory SSD B iodrive Device iodrive Device Default Settings 274.17 633.07 742.28 63% 15% BUFSIZE 512k BUFNO 4k NOSGIO 185.93 208.83 421.85 56% 50% BUFSIZE 256k BUFNO 1k SGIO 287.03 346.28 505.34 43% 31% BUFSIZE 512k BUFNO 4k SGIO 275.41 322.12 599.51 54% 46% 2x Concurrent Programs BUFSIZE 512k BUFNO 4k NOSGIO 330.42 544.42 825.47 60% 34% 2x Concurrent Programs BUFSIZE 512k BUFNO 4k SGIO 319.87 381.12 862.56 63% 56% Total real time 1,673.1s 2,435.84 s 3,957.01s 58% 38% Table 1: Summary of Performance with a 10GB input SAS data set 4
Figure 1 below illustrates the data transfer for each storage device, with the server running two concurrent SAS programs with BUFSIZE 512k, BUFNO 4k, and SGIO enabled. This pattern is typical of the results observed. Fusion iomemory iodrive Figure 1: Data Transfer (GB/s) by device Amadeus Software s performance monitoring found that the Fusion iomemory iodrive devices consistently achieved a higher I/O throughput when compared to the other devices. The higher I/O leads to a higher CPU utilization and a shorter elapsed processing time. Further Selected Findings Bandwidth Testing The maximum observed bandwidth during the testing was on the Fusion iomemory iodrive device, which achieved almost 4GB/s with five concurrent SAS sessions on the server (and with each SAS session using a 10GB input data set). Solid State Technology B and hard drives peaked at 2.5GB/s and 2GB/s, respectively. Additionally, the Fusion iomemory iodrive device achieved high bandwidth consistently, whereas the SSD array and hard disk only achieved their highs in bursts. This pattern is clear in Figure 1. Optimal Configuration for Data Sets Over 2GB Adjusting the BUFNO and BUFSIZE options improved completion times for SAS data sets larger than 2GB. However, it is worth noting that larger SAS data sets require more memory to complete a step in the program. We have found that switching on SGIO significantly reduces the required memory. 5
Optimal Settings for Single Program Runs When executing a single program, the best results were found with the following settings: SGIO: OFF BUFSIZE: 512K BUFNO: 4K However, as noted above, these settings required a very large amount of memory. Optimal Setting for Parallel Program Runs When executing several SAS sessions in parallel, the best results were found with the following settings: SGIO: ON BUFSIZE: 512K BUFNO: 4K It is worth noting that with SGIO switched off, Amadeus Software exhausted the memory available to the SAS sessions. This is one reason why Amadeus recommends regular performance monitoring to ensure that appropriate SAS system options and values are used in SAS environments. When five programs, each using 10GB input data sets, ran in parallel execution, the Fusion iomemory iodrive device performed with 30 percent more efficiency than the Solid-state Technology B. The mechanical drive was unable to complete the tests in this configuration due to its low capacity. When running tests in this configuration, Amadeus Software observed that the Fusion iomemory iodrive device also achieved a maximum throughput of 4GB/s compared Solid State Technology B, which peaked at 0.8GB/s. With the Fusion iomemory iodrive device, Amadeus observed individual procedure steps peaking at a five-fold performance gain when compared to the mechanical disk. CONCLUSION The results clearly show that the Fusion iomemory iodrive device is the most efficient solid-state technology, and significantly outperforms mechanical disks and alternative solid-state devices. Amadeus Software found the Fusion iomemory iodrive device reduced job completion time by approximately two-thirds when concurrent SAS sessions were processing in a well- configured SAS environment. On average (including environments with single SAS sessions), Amadeus Software observed a 58 percent average decrease in required time to complete an SAS job. The return on investment for a Fusion iomemory iodrive device can be measured as follows: Creating a more powerful SAS system that can complete more jobs and more complex jobs, faster. Reducing the size of physical infrastructure. Submitting more jobs to each server in parallel increases the workload and utilisation of SAS servers. In addition, environmental returns can be measured. 6
Buying less storage. Fusion iomemory iodrive devices can store several terabytes of storage in each server. This makes it possible to eliminate the physical devices in the server room required with NAS and SAN arrays. Reclaiming space and storage, and reducing energy costs. By moving active data into SAS servers, SAN and NAS arrays are freed for other functions, less space is needed for shelves of striped disk arrays, and less energy is needed to power and cool them. The performance results discussed herein are based on internal testing and use of the referenced products. Results and performance may vary according to configurations and systems, including drive capacity, system architecture, and applications. 2014 SanDisk Corporation. All rights reserved. SanDisk is a trademark of SanDisk Corporation, registered in the United States and other countries. Fusion iomemory, iodrive and others are trademarks of SanDisk Enterprise IP LLC. Other brand names mentioned herein are for identification purposes only and may be the trademarks of their respective holder(s). 7