The linked image cannt be displayed. The file may have been mved, renamed, r deleted. Verify that the link pints t the crrect file and lcatin. Technical Brief Caching Sftware Perfrmance Test: Micrsft SQL Server Acceleratin with FlashSft Sftware 3.8 fr Windws Server SanDisk Crpratin Crprate Headquarters 951 SanDisk Drive, Milpitas, CA 95035 www.sandisk.cm
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Table f Cntents Intrductin... 3 System Under Test... 4 Hardware... 4 Sftware... 5 Test Prcedure... 5 General Setup (fr all tests)... 6 Bare Metal Wrklad Acceleratin Test Setup... 6 File Exclusin Test Setup... 8 Measured Results... 10 Wrklad Acceleratin Test... 10 Perfrmance Chart... 10 System Activity... 11 File Exclusin Test... 12 Perfrmance Chart... 13 Analysis / Cnclusin... 13 2
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Intrductin The bjective f this paper is t present the perfrmance gains that can be achieved n bare metal and virtualized cmputing envirnments thrugh the use f a hst- based slid- state strage cache enabled by the FlashSft sftware frm SanDisk. Micrsft SQL Server 2014 was installed n a server running Micrsft Windws Server 2012 R2 and perfrmance was measured frm a client cmputer using HammerDB. The database fr the test was cntained n a strage array cmprising SAS- attached hard disk drives (HDDs), which was attached directly t the hst server. 3
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB System Under Test The cmputers used fr this test were cnfigured as fllws: Hardware Hst Server: Dell PwerEdge R730xd CPU: 2x Intel Xen E5-2690 v3 @ 2.60GHz, 12 cres each System Memry (RAM): 192GB DDR3 Strage Cntrller: Dell PERC H730 Mini RAID cntrller (embedded); cache size: 1024MB! HDD Strage: 12x 300GB SAS 6Gb/s 10K RPM, Seagate ST300MM0006! Flash Memry (SSD): 1x 1.3TB PCIe Fusin imemry SX350 1300 (cache) Direct Attached Strage Array: Dell PwerVault MD3200 10x 600GB SAS 6Gb/s 10K RPM, Tshiba AL13SEB600 The drives were frmatted int tw virtual disks: DB (2.8TB) and LOG (1.1TB) Directly cnnected t hst server via SFF- 8088 Mini SAS cable Client Cmputer: Dell PwerEdge R720xd CPU: 2x Intel Xen E5-2603 @ 1.80GHz, 4 cres each System Memry (RAM): 96GB DDR3 Strage Cntrller: Dell PERC H310 Mini RAID cntrller (embedded)! HDD Strage: 2x 500GB SAS 6Gb/s 7.2K RPM, Seagate ST500NM0001 Figure 1: Diagram f hardware cnfiguratin fr tests 4
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Sftware Hst Server Sftware Windws Server 2012 R2, Windws Server 2012 R2 Hyper- V Micrsft SQL Server 2014! Micrsft SQL Server Management Studi: 12.0.4100.1! Micrsft Analysis Services Client Tls: 12.0.4100.1! Micrsft Data Access Cmpnents (MDAC): 6.3.9600.16384! Micrsft MSXML: 3.0.6.0! Micrsft.NET Framewrk: 4.0.30319.33440! Buffer pl size / RAM reservatin fr SQL Server: 16GB Client Cmputer Sftware Windws Server 2012 R2 HammerDB: 2.18 Caching Sftware FlashSft fr Windws Server (versin 3.8.0)! fidrv Driver Versin: 4.1.0.62685 Flash Memry Driver: Fusin imemry SX350-1300, VSL Driver 4.2.1 Test Prcedure The benchmark test was cnducted multiple times t measure and cmpare perfrmance f the nn- accelerated HDD strage backend and the same HDD backend accelerated using a FlashSft sftware write back cache. Tw sets f tests were perfrmed: Wrklad Acceleratin: Measure the increased number f Transactins Per Minute (TPM) pssible fr the same strage cnfiguratin using a FlashSft write- back cache cmpared t the nn- accelerated HDD baseline. File Exclusin: Observe the increased number f Transactins Per Minute pssible using the same FlashSft accelerated wrklad n a virtual machine with certain files (perating system and lgs) excluded frm acceleratin. 5
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB General Setup (fr all tests) 1) Install and cnfigure Windws Server 2012 R2 and SQL Server 2014 n the hst server. 2) Install and cnfigure flash memry (SSD) and FlashSft sftware n the hst server. 3) Install and cnfigure HammerDB n the client cmputer. 4) Cnnect the hst server t the backend strage array using a SAS cable. Bare Metal Wrklad Acceleratin Test Setup 1) On the hst server map t the MD strage array and create addressable strage vlumes fr the benchmark test: Data ( F: ) 488GB (RAID 5) Lg ( G: ) 195GB (RAID 10) Cache ( H: ) 135GB (managed by Fusin imemry VSL driver 4.2.1) Any additinal capacity shuld remain unallcated Figure 2: Strage vlume mappings n the hst server 6
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB 2) Cnfigure HammerDB n the client cmputer using the fllwing settings: Database size: 409GB ODBC Driver: ODBC Driver 11 fr SQL Server Ttal Transactins per User: 1,000,000,000 Ramp up Time: 150 minutes Test Duratin: 60 minutes Figure 3: HammerDB client settings 3) Run HammerDB t measure perfrmance f the nn- accelerated backend (baseline). 4) Enable FlashSft sftware t accelerate the DATA disk nly (d nt accelerate the OS r LOG disks). 5) Run HammerDB t measure perfrmance f the accelerated backend. 6) Cmpare results. 7
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB File Exclusin Test Setup Starting with the Wrklad Acceleratin Test cnfiguratin described abve: 1) Recnfigure strage vlumes f the MD strage array as fllws: Data ( F: ) 2,048GB (RAID 5) Lg ( G: ) 195GB (RAID 10) Cache ( H: ) 102GB (managed by Fusin imemry VSL driver 4.2.1) Any additinal capacity shuld remain unallcated Figure 4: Strage vlume mappings n the hst server 2) Enable Hyper- V n the hst server and create ne virtual machine. a. RAM: 39GB b. Create the fllwing VHDX strage vlumes in the virtual machine. Their assciated disk image files shuld be stred n vlume F: f Disk 0 n the physical hst server. i. Data: ( F: ) 300GB (295GB) VHDX ii. Lg: ( G: ) 100GB VHDX iii. OS: ( C: ) 100GB VHDX 8
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Figure 5: Strage vlumes f the virtual machine Figure 6: Disk image files n the hst server c. Install and cnfigure Windws Server 2012 R2 as the guest perating system n the virtual machine. d. Install and cnfigure SQL Server 2014 n the virtual machine. 3) Enable FlashSft n the hst server t accelerate the disk image files lcated n vlume F: which in turn will accelerate perfrmance f the Data, Lg and OS VHXD vlumes f the virtual machine. 4) Run HammerDB n the client cmputer t measure SQL perfrmance f the accelerated virtual machine. 5) On the hst server disable FlashSft acceleratin f the disk image files fr Lg and OS. 6) Run HammerDB n the client cmputer t measure SQL perfrmance f the virtual machine with nly the Data vlume accelerated. 7) Cmpare results. 9
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Measured Results Wrklad Acceleratin Test Wrklad Acceleratin Test (Bare Metal) Test Baseline Strage backend nt accelerated. Accelerated Data files accelerated with FlashSft write- back cache. Summary HammerDB benchmark Database size: 409GB Ramp up time: 150 minutes Test time: 60 minutes SQL buffer pl size: 16GB HammerDB benchmark Database size: 409GB Cache size: 135GB (33% size f wrklad) Write- back cache OS and lg nt accelerated Ramp up time: 150 minutes Test time: 60 minutes SQL buffer pl size: 16GB Scre (Transactins Per Minute) 93,042 TPM 403,058 TPM x4.38 Imprvement ver Baseline Table 1: Wrklad acceleratin test Perfrmance Chart Bare Metal Perfrmance Increase with FlashSj Sjware Transacins Per Minute 450000 400000 350000 300000 250000 200000 150000 100000 50000 0 93,042 TPM Baseline 403,058 TPM Accelerated Figure 7: Increased TPM perfrmance with FlashSft write- back caching 10
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB System Activity Figure 8: Baseline SQL Transactins Figure 9: Accelerated SQL Transactins Figure 10: Baseline disk activity Figure 11: Accelerated disk activity 11
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB File Exclusin Test File Exclusin Test Benchmark (Hyper- V) Baseline Test Strage backend nt accelerated. Accelerated w/ File Exclusin All files accelerated with FlashSft write- back cache. Accelerated w/ File Exclusin Only database files accelerated with FlashSft write- back cache. OS and lg VHDx files excluded frm acceleratin. Table 2: File exclusin test Summary HammerDB benchmark Database size: 295GB Ramp up time: 150 minutes Test time: 60 minutes SQL buffer pl size: 16GB HammerDB benchmark OS, lg and database accelerated Database size: 295GB Cache size: 135GB (46% size f wrklad) Write- back cache Ramp up time: 150 minutes Test time: 60 minutes SQL buffer pl size: 16GB HammerDB benchmark Database accelerated nly (lg, OS excluded) Database size: 295GB Cache size: 135GB (46% size f wrklad) Write- back cache Ramp up time: 150 minutes Test time: 60 minutes SQL buffer pl size: 16GB Scre (Transactins Per Minute) 71,744 TPM 206,852 TPM x2.88 Imprvement ver Baseline 425,527 TPM x5.93 Imprvement ver Baseline 12
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB Perfrmance Chart File Exclusin Perfrmance Increase with FlashSj Sjware 450000 425,527 TPM Transacins Per Minute 400000 350000 300000 250000 200000 150000 100000 50000 71,744 TPM 206,852 TPM 0 Baseline Accelerated (All) Accelerated (File Exclusin) Figure 12: Enhanced caching perfrmance utilizing file- exclusin feature Analysis / Cnclusin Data measured in the tests illustrate hw FlashSft sftware can significantly imprve perfrmance f Micrsft SQL Server wrklads n bth bare metal servers and Hyper- V virtual machines cmpared t traditinal all- HDD backend strage. The tests were cnducted directly n the bare metal server and virtualized with Hyper- V using a Micrsft SQL Server wrklad measured by the HammerDB benchmark tl. Althugh the tests were cnstructed t be cmparable t the cnditins typically encuntered in real- wrld cmputing and t generate data that reveal representative applicatin perfrmance, the tests can nly be cnsidered a demnstratin f the capability f FlashSft sftware the data shuld nt be interpreted as the perfrmance impact f FlashSft sftware fr all wrklad types and strage envirnments. The actual perfrmance f any caching slutin is highly dependent upn the wrklad and the cmputing envirnment in which it is used. 13
Caching Sftware Perfrmance: FlashSft 3.8 fr Windws Server Micrsft SQL Server and HammerDB The applicatin perfrmance test cnducted n the bare metal hst server demnstrates a 4.38 times increase in transactins per secnd as measured by HammerDB. The cache was nly 1/3 the size f the database (wrklad) and was cnfigured in write- back cache mde. The system activity graphs (figures 8 thrugh 12) shw peratin f the cache as it warms up and enhances perfrmance, as shwn by the increased levels and intensity f SQL transactins and disk activity. The file exclusin test highlights peratin f a unique FlashSft feature that allws the administratr t exclude individual files that wuld nt typically benefit frm caching. This frees caching resurces t be applied tward cacheable files such as the database instead f the lg r perating system. Additinally, it eliminates thrashing f the cache that ccurs as the lg is regularly updated with nn- cacheable data intermixed with data f the database. The file exclusin test was cnducted within a virtual machine t allw the lg and perating system VHDx files t be excluded frm cache acceleratin. The parameters f the database and virtual machine strage were adjusted fr this test t perate in the virtual machine. Table 2 illustrates a 2.8 times increase f perfrmance ver baseline when the lg and perating system files are accelerated tgether with the database. Excluding the lg and perating system in rder t accelerate the database nly further dubled the accelerated perfrmance, resulting in nearly 6 times increase in perfrmance cmpared t the baseline. Specificatins are subject t change. 2015 SanDisk Crpratin. All rights reserved. SanDisk is a trademark f SanDisk Crpratin, registered in the U.S. and ther cuntries. FlashSft and Fusin imemry are trademarks f SanDisk Crpratin. Other brand names mentined herein are fr identificatin purpses nly and may be the trademarks f their respective hlder(s). 5033EN 20151020 Rev. 1 14