Proving out flash storage array performance using swingbench and slob
Download as PPTX, PDF1 like440 views
This document discusses testing the performance of a flash storage array using the tools Swingbench and SLOB. It provides details on running tests with SLOB to measure IOPS and latency for random reads and writes. It also describes using Swingbench to test throughput by running the Sales History benchmark against a 500GB schema, varying configuration settings like parallelism and indexes. The results of these tests are analyzed to demonstrate the performance of the flash storage array.
Proving out flash storage array performance using swingbench and slob
- 1. Proving out Flash Storage Array Performance using Swingbench and SLOB Kapil Goyal, Fidelity Investments @kapil_goyal
- 2. Who am I? Lead for Oracle database performance team at Fidelity Investments Worked with Oracle Consulting and got exposure to many different companies' database performance challenges Speaker at ECO, BIWA, DOUG, Hostos, Enkitec Expo, IOUG and AIOUG Published articles for DOUG and IOUG SELECT magazine Certified OCP DBA 8i to 11g
- 3. Storage Fundamentals • IOPS • Throughput • Latency • SAN connectivity • SAN Cache
- 4. IOPS • # of distinct Input/Output operations (reads or writes) per second • Several vendors advertise "just" higher IOPS, sort of meaning less without other metrics like • Latency • Block Size - 2k, 4k, 8k, 16k, 32k...? • Read write ratio
- 5. Latency • A measure of the time required for a sub-system or a component in that sub-system to process a single storage transaction or data request • Latency matters!! • Varies by block size
- 6. Latency • DB level latency can be in sub ms still with spinning drives Reference - http://carymillsap.blogspot.com/2009/02/dang-it-people-theyre-syscalls-not.html • Wait event is SYSCALL • Db file sequential read – pread call (linux) • wait n. In an Oracle context, the approximate response time of what is usually a single operating system call (syscall) executed by an Oracle kernel process. • SAN Cache • FS Cache • Oracle data doesn’t show IO queue at OS level • Oracle has no visibility if IO got serviced by FS cache, SAN cache or by spinning drive
- 7. Throughput • How much data can be transferred from one location to another in a given amount of time Or • MegaBytes/Second • Important for large scans/sequential reads
- 8. SAN connectivity fundamentals • Zoning • Port base • WWN base
- 9. Does your Database need Flash Storage? • Is your database IO sensitive? • Does it need consistent and low latency? • Does it currently have high latency and driving a lot of IO? • Database server CPU starved? • 3 offerings depending upon workload: • Hybrid Storage with spinning drives with Cache (RAM or FLASH) • All Flash arrays • Exadata
- 10. AFA (All Flash Array) POC – Choose the right tool Reference - https://www.pythian.com/blog/oracle-database-load-testing-tools-overview/ • Tools available: • Orion (Oracle IO numbers) • Winsat • Iometer • HammerDB • Oracle RAT (Real Application Testing) • LoadRunner • SLOB • Swingbench • .. • ..
- 11. • SLOB – For IOPS (To Test Physical IO) • Swingbench • Order Entry - introduces heavy contention on a small number of tables and is designed to stress interconnects and memory • Sales History - to test the performance of complicated queries when run against large tables. Tools we picked
- 12. • SLOB – For IOPS (To Test Physical IO) • Swingbench • Order Entry - introduces heavy contention on a small number of tables and is designed to stress interconnects and memory • Sales History - to test the performance of complicated queries when run against large tables. Tools we picked
- 13. • 2 Linux physical servers: • Dell PowerEdge R630 – Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz (Haswell) – 20 Cores – 378G RAM, RAM Speed - DDR4 384G (24x16g dimms) @1866@MHz – HBA – QLogic QLE2662 16Gb FC Adapter • What max throughput we can drive on each box – ~3200 MBPS • 2 node RAC ASM 12.1 Database Oracle Parameter Swingbench SLOB SGA_TARGET 15G 2G SGA_MAX_SIZE 15G 15G PGA_AGGREGATE_TARGET 100G 100G PGA_AGGREGATE_LIMIT 200G 200G JOB_QUEUE_PROCESSES 80 80 PARALLEL_MAX_SERVERS 552 552 PARALLEL_MIN_SERVERS 288 288 REDO CONFIGURATION 10 Groups(min). 1GB each. 10 Groups(min). 1GB each. TEMP TABLESPACE 200GB 200GB Environment Details
- 14. SLOB Reference- https://kevinclosson.net/slob/ • IO workload generation toolkit • Supports following testing • Oracle Logical Read scaling • Physical random single block read (db file sequential read wait event) • Random single block writes (DBWR flushing capacity) • Redo logging IO • Best tool to test IO subsystem
- 15. SLOB • Create ~500GB dataset (256 SLOB schemas each 2 GB) • Ran several tests with different # of users - 2,4,8,16,32,64,92,128,160,192,224 and 256
- 16. Initial Setup https://kevinclosson.net/2014/08/04/slob-deployment-a-picture-tutorial/ • Create a tablespace for SLOB • Run setup.sh - To load SLOB schemas • Execute runit.sh - to run real tests
- 17. Slob.conf UPDATE_PCT= 0 to 50 -–Update Activity RUN_TIME=1800 --seconds WORK_LOOP=0 SCALE=256000 --Numbers of Blocks WORK_UNIT=64 REDO_STRESS=HEAVY #LITE LOAD_PARALLEL_DEGREE=4 # Settings for SQL*Net connectivity: ADMIN_SQLNET_SERVICE=slob_poc5 SQLNET_SERVICE_BASE=slob_poc5 #SQLNET_SERVICE_MAX=2 #SYSDBA_PASSWD=xxxxx SYSDBA_PASSWD=manager
- 18. SLOB Test Setup oracle@hostname:poc51> srvctl status service -d poc5_wlk Service slob_poc5 is running on instance(s) poc51,poc52 Service slob_poc51 is running on instance(s) poc51 Service slob_poc52 is running on instance(s) poc52 tnsnames.ora slob_poc51 = (DESCRIPTION = (ADDRESS = (PROTOCOL = TCP)(HOST =xxxx)(PORT = 1521)) (CONNECT_DATA = (SERVER = DEDICATED) (SERVICE_NAME = slob_poc51))) slob_poc52 = (DESCRIPTION = (ADDRESS = (PROTOCOL = TCP)(HOST = xxxx)(PORT = 1521)) (CONNECT_DATA = (SERVER = DEDICATED) (SERVICE_NAME = slob_poc52))) slob_poc5 = (DESCRIPTION = (ADDRESS = (PROTOCOL = TCP)(HOST = xxxx)(PORT = 1521)) (CONNECT_DATA = (SERVER = DEDICATED) (SERVICE_NAME = slob_poc5))) Test it sqlplus user1/user1@slob1 <<EOF show parameter instance_name EOF sqlplus user1/user1@slob2 <<EOF show parameter instance_name EOF
- 19. Some underbar parameter simple.ora # The following are needed to force db file sequential read. Omit for direct path short scans. # If you would like to investigate a blend of short scans mixed with single-block random I/O # then comment out these three parameters : _db_block_prefetch_limit = 0 _db_block_prefetch_quota = 0 _db_file_noncontig_mblock_read_count = 0
- 20. SLOB Caveat • It overwrites logs • Best for IOPS test but not good for driving high throughput
- 22. Runit wrapper script Better naming convention Doesn’t overwrite files Make Buffer Cache small to stress IO subsystem
- 23. IOPS/MBPS script Number of Number of Total Number of Number of Total Total Total DATE_TIME DATE_TIME Small reads Small writes Small IOPS large reads large writes Large IOPS MBPS read MBPS written Total MBPS Total IOPS -------------------- -------------------- ------------ ------------- ----------- ------------ ------------- ----------- ---------- ------------- ---------- ---------- 02/05/18_00_19_00_21 02/05/18_00_19_00_21 222 2716 2938 18 14 32 16 35 51 2970 02/05/18_00_21_00_51 02/05/18_00_21_00_51 70374 32138 102512 222 694 916 769 931 1700 103429 .. .. 02/05/18_07_51_08_56 02/05/18_07_51_08_56 18 3 21 1738 10 1748 1728 2 1730 1769 Throughput = IOPS x I/O size In what case IOPS and Throughput could be same?
- 24. Other tools – db_capacity.sql • Leverages google chart • Thanks to Mauro and Carlos for help
- 25. Other tools – db_capacity.sql
- 26. Other tools – db_capacity.sql
- 27. Other tools – db_capacity.sql
- 28. Update% - 10 (To change read/write ratio)
- 29. Update% - 10 (To change read/write ratio) 0 1000 2000 3000 4000 2 4 8 16 32 64 96 128 224 256 # of SLOB Sessions MBPS MBPS 0 100000 200000 300000 400000 2 4 8 16 32 64 96 128 224 256 # of SLOB Sessions IOPS IOPS 0 1 2 3 4 5 2 4 8 16 32 64 96 128 224 256 ms # of SLOB Sessions Single Block Read Latency (ms) Avg Max 0 2 4 6 8 2 4 8 16 32 64 96 128 224 256 ms # of SLOB Sessions log file parallel write (ms) Avg Max
- 30. Update% - 10 (To change read/write ratio) 27082 49337 90660 163114 292842 347478 345245 323205 326606 325784 0.39 0.43 0.47 0.51 0.56 1.02 1.57 2.23 3.62 4.04 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 50000 100000 150000 200000 250000 300000 350000 400000 2 4 8 16 32 64 96 128 224 256 Rd RdLatency(ms) IOPS # of SLOB Sessions IOPS Rd Latency(ms)
- 31. Swingbench testing (Throughput) Shwizard to build Sales History Schema 500GB schema was created (with range partition) Several tests (SH) were run with different parallel/Index settings Swingbench config info • NumberOfUsers 16 • MinDelay 0 • MaxDelay 0 • InterMinDelay 1000 ms • InterMaxDelay 2000 ms • RunTime 1h:5 min • StatsCollectionStart 0:4 min • StatsCollectionEnd 1h:4 min http://dominicgiles.com/swingbench.html
- 32. Xml config file <?xml version="1.0" encoding="UTF-8" standalone="yes"?> <SwingBenchConfiguration xmlns="http://www.dominicgiles.com/swingbench/config"> <Name>Sales History</Name> <Comment>Configuration file for the Sales History Benchmark</Comment> <Connection> <UserName>sh_500</UserName> <Password>sh_500</Password> <ConnectString>//eihtlrac-scan/slob_poc5</ConnectString> <DriverType>Oracle jdbc Driver</DriverType> <Properties> <Property Key="FetchSize">20</Property> <Property Key="StatementCaching">50</Property> <Property Key="TcpNoDelay">true</Property> </Properties> <InitilizationCommands> <Command>alter session set sql_trace=false</Command> </InitilizationCommands> </Connection> <Load> <NumberOfUsers>16</NumberOfUsers> <MinDelay>0</MinDelay> <MaxDelay>0</MaxDelay> <InterMinDelay>1000</InterMinDelay> <InterMaxDelay>2000</InterMaxDelay> <QueryTimeout>55</QueryTimeout> <MaxTransactions>-1</MaxTransactions> <RunTime>1:5</RunTime>
- 33. Charbench DIR='/u01/swingbench' BDT=$(date +%Y%m%d) BTM=$(date +%H%M%S) $DIR/bin/charbench -c $DIR/configs/ei_sh.xml -r ${FILE_NM}_results_${BDT}_${BTM}.xml
- 34. Output xml <?xml version="1.0" encoding="UTF-8" standalone="no"?> <Results xmlns="http://www.dominicgiles.com/swingbench"> <Overview> <BenchmarkName>Sales History</BenchmarkName> <Comment>Configuration file for the Sales History Benchmark</Comment> <TimeOfRun>Feb 7, 2018 9:47:44 PM</TimeOfRun> <TotalRunTime>1:00:00</TotalRunTime> <TotalLogonTime>0:00:00</TotalLogonTime> <TotalCompletedTransactions>40</TotalCompletedTransactions> <TotalFailedTransactions>1</TotalFailedTransactions> <AverageTransactionsPerSecond>0.01</AverageTransactionsPerSecond> <MaximumTransactionRate>5</MaximumTransactionRate> <AverageResponse>1300652.25</AverageResponse> <MinimumResponse>632</MinimumResponse> <MaximumResponse>3102874</MaximumResponse>
- 35. Test Summary
- 36. Vendor rating • Support • Security • Reporting • Reliability • Performance • Management • Functionality • Cost • Storage efficiency – (Dedup, compression) • Scalability
- 37. Appendix - A Test case – "Spectre" and "Meltdown" vulnerability patching with BIOS update
- 39. Appendix-C Stack Ranking of Oracle databases
- 40. Stack rank your databases Database Rank IOPS GB/s Latency Cores Storage ORAprod 1 1.5M 100 <1ms 128 100TB ORAprod 2 1.4M 47 <1ms 128 75TB DB2prod 3 1.3M 46 <1ms 128 75TB ORAprod . . . . . . DB2prod . . . . . . SQLprod . . . . . . ORAprod . . . . . . DB2prod . . . . . . SQLprod . . . . . . ORAtest . . . . . . DB2test . . . . . . SQLtest . . . . . . ORAdev . . . . . . SQLprod . . . . . . SQLtest . . . . . . SQLdev 3000+ 100 1 >10ms 2 20GB High End Physical High End Virtual Virtual On-Prem As-A-Service Off-Prem As-A-Service 80% virtual start here
- 41. Enterprise wide 95th, 98th and 99th database level stats
- 43. Q&A
Editor's Notes
- #4: As a DBA it is very important to understand these basic.
- #7: DB was on FS good response time, moved to ASM and horrible IO response time. SGA was too small
- #9: Dual fabric – A and B,
- #10: Earlier we had only spinning drives with lots of mix workload and inconsistent latency or Exa, we needed something in-between.
- #11: Invited 5 vendors, put together a matrix what they have to offer in terms of IOPS, Throughput & latency
- #30: Good to plot db file parallel write as well (Writes from Buffer Cache)
- #31: Another way to compare IOPS and Read Latency