11Performance- and Power-Efficient5SSDs120xHDDs-36,00010PS120xSSDs-12GB/sec490FC15KRPMdrives- 4,200,00010PS1,452watts- 36GB/sec8.8TB288wattsPerHDD3.8TB:100MB/sec(R/W)PerSSD:3001OPS250/170MB/sec（R/W）12.1watts(active).35,0001OPS(Read)2.4watts(activeReadthroughput(IOPS)BandwidthPowerCost115x3x5xAdaptedfromFrankHardy@Intel03/07/2009talk,FC=fiberchannel)
Performance- and Power-Efficient • 120x SSDs – 4,200,000 IOPS – 36GB/sec – 288 watts – 3.8TB – Per SSD • 250/170MB/sec (R/W) • 35,000 IOPS (Read) • 2.4 watts (active) 11 Read throughput (IOPS) 115x Bandwidth 3x Power Cost 5x * Adapted from Frank Hardy@Intel 03/07/2009 talk, FC=fiber channel) • 120x HDDs – 36,000 IOPS – 12GB/sec – 1,452 watts – 8.8TB – Per HDD • 100MB/sec (R/W) • 300 IOPS • 12.1 watts (active)
12Challenge 1: AffordabilityA full-ssD based storage solution- Example: Gordon HPC Cluster*.Data-centricScientificApp00·64TBDRAM+300TBFlash· $20,000,000 funding from NSFSDSC· 3-4 years lifespan$5,000,000/year- In reality·Highperformance comesfrom veryhighcost·Notaffordablefor mostdatacenters*source:http://www.internetnews.com/hardware/article.php/3847456
Challenge 1: Affordability • A full-SSD based storage solution – Example: Gordon HPC Cluster* • Data-centric Scientific App. • 64TB DRAM + 300TB Flash • $20,000,000 funding from NSF • 3-4 years lifespan – In reality • High performance comes from very high cost • Not affordable for most data centers 12 *source: http://www.internetnews.com/hardware/article.php/3847456 $5,000,000/year
13Challenge 2: Performance DynamicsRandomread4KBinthe1024MBspacewith1~321/Ojobs(differentdataallocationsamongflashchipsresultindifferentperformance)PerformanceDynamicsonIntelX25-ESSD3.53,000μsAveragelatency(msec)he worstcase2.54.2x higher()eLatencyThebest caseLSBest CaseQueueDepth（#of*700μsconcurrent1/oJobs)0.5151617181920212223242526272829303132QueueDepth*DellPrecisionT3400,Intelcore2Duo2.66Ghz,4GBMemory.IntelX25-ESSD,FC9,Linuxkernel26.27,PostgreSQL8.3.4
0 0.5 1 1.5 2 2.5 3 3.5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 Latency (msec) Queue Depth Performance Dynamics on Intel X25-E SSD Worst Case Best Case Challenge 2: Performance Dynamics • Random read 4KB in the 1024MB space with 1~32 I/O jobs (different data allocations among flash chips result in different performance) 13 4.2x higher Latency * Dell Precision T3400, Intel core2Duo 2.66Ghz, 4GB Memory, Intel X25-E SSD, FC 9, Linux kernel 26.27, PostgreSQL 8.3.4 The worst case The best case Average latency (msec) Queue Depth (# of concurrent I/O Jobs) 3,000µs 700µs
14Challenge3:ResourceunderutilizationDatabasequeryexecutionsSSD-opt.25-ESSDStarSchemaBenchmark(SsB)Qu(127.2x)140However,thehighperformancepotentialofSsDcannotbeautomaticallytappedwithoutextensiveresearcheffortsSSD-baseline80X)dnpaadsDisk(27.5x)SSD-Baseline60SSD-OptimizedDisk(21min)4020Q1.1Q1.2Q2.2Q3.2Q4.3*DellPrecisionT3400,Intelcore2Duo2.66Ghz4GBMemory,IntelX25-ESSD,FC9,Linuxkernel2.6.27,PostgreSQL8.3.4
• Database query executions 0 20 40 60 80 100 120 140 Q1.1 Q1.2 Q2.2. Q3.2 Q4.3 Speedup (X) Star Schema Benchmark (SSB) Queries on Intel X25-E SSD Disk SSD-Baseline SSD-Optimized Challenge 3: Resource underutilization 14 * Dell Precision T3400, Intel core2Duo 2.66Ghz, 4GB Memory, Intel X25-E SSD, FC 9, Linux kernel 2.6.27, PostgreSQL 8.3.4 SSD-opt. (127.2x) Disk (21min) SSD-baseline (27.5x) 79% of the performance potential is not utilized The peak performance of SSD can revolutionize the existing storage systems. However, the high performance potential of SSD cannot be automatically tapped without extensive research efforts
15Challenge 4: Reliability of SSDProgram-erasecycles (P/Ecycles):Mostcommerciallyavailableflashproductsareguaranteedtowithstandaround100,000P/Ecycles.-Inreality,theP/EcycleswilldecreaseforlowendSsDproductssothattheprice can be affordable.- Wear leveraging: dynamically remapping blocks in order to spreadwriteoperationsbetweensectors.ReadDisturbCauseothercellsnearthebeingreadtochangeovertimeiftheyarenotrewritten.LimitedCapacity- Trade-off between large capacityand reliability subjecttoacost.- A large portion of the space is used for over-provisioning
Challenge 4: Reliability of SSD • Program-erase cycles (P/E cycles): – Most commercially available flash products are guaranteed to withstand around 100,000 P/E cycles. – In reality, the P/E cycles will decrease for low end SSD products so that the price can be affordable. – Wear leveraging: dynamically remapping blocks in order to spread write operations between sectors. • Read Disturb – Cause other cells near the being read to change over time if they are not rewritten. • Limited Capacity – Trade-off between large capacity and reliability subject to a cost. – A large portion of the space is used for over-provisioning. 15