NVRAMOS ‘11 November 8, 2011, Jeju SSD Characterization: From Energy Consumption's Perspective Youjip Won Hanyang University Outline p Motivation p Related Works p SSD Organization and Energy Consumption p Channels, Ways and Clusters p Trivia in Measurement Methodology p Case Study p Power Budget p Conclusion Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 2/68 Motivation 3/68 Motivation Understanding of Internal Mechanism of Storage Device is very important! p Hard disk w Sector layout: cylinder serpentine vs. surface serpentine vs. hybrid serpentine w Number of zones w Degree of track skew w Disk scheduling algorithm Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 4/68 Characterizing HDD: Sector Layout p Jongmin Gim et al, ACM ToS 6, 2 (July 2010) Traditional Spindle Spindle Cylinder serpentine Spindle Youjip Won Surface serpentine Zone rewind Spindle Hanyang University NVRAMOS’11, Jeju, Korea 5/68 Characterizing Storage Devices HDD Characterization is via measuring Seek time and Rotational Latency. Characterizing SSD… what do we use?... Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 6/68 Motivation How do we figure out the internals of SSD? p What is available w the number of channels w the number of chips/packages per channels p What is not available? w Sector placement , Garbage collection algorithm Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 7/68 Operations on NAND Flash Cell program erase 18~35V Control Gate Control Gate Oxide Layer Floating - - Gate Source Youjip Won Floating Gate Drain Source - - - Drain Substrate Substrate Programed: 1 -> 0 Erased : 0 -> 1 Hanyang University NVRAMOS’11, Jeju, Korea 8/68 Operations on NAND Flash Cell Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 9/68 SSD Characterization We will use “Energy Consumption” Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 10/68 Related Works 11/68 Related Works p Dongkun Shin et al, NVRAMOS 2010 Spring Identifying the relationship between workload characteristics and aggregate power consumption for each workload Applied mixed workloads(random, sequential, etc) to SSDs with different request sizes and varied the file systems Measured the Power Consumption(measured voltage change) p Laura M. Grupp et al, MICRO 2009 Read Power Consumption of Flash Memory Basic Operations Program Erase Custom Board Basic Operation + Flash Memory Youjip Won Hanyang University Examine the average power consumption(W) and energy(J) NVRAMOS’11, Jeju, Korea 12/68 Related Works p Euiseong Seo et al, HotPower’08 Analyzed the power consumption patterns of the SSDs different hardware configurations to the various atomic operations and the combination of file systems and workloads. p Vidyabhushan Mohan et al, Date ’10 CACTI 5.3 developed a detailed power model for the NAND flash chip itself with CACTI 5.3 Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 13/68 SSD Organization and Energy Consumption 14/68 Power Consumption of Storage Devices: System Boot Active current of HDD stand-by current s SSDs have various stand-by currents Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 15/68 SSD Organization DRAM Controller Bus Host I/O Physical Interface … … … … … … … … … … … … … … … … Host Interface NAND Flash Memory Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 16/68 Characteristics of SSD Behavior =Page Write =Write Complete Youjip Won … … … … … … … … … … … … … … … … Hanyang University NVRAMOS’11, Jeju, Korea 17/68 IO Size vs. Energy Consumption Intel X25M Youjip Won … … … … … … … … … … … … … … … … … … … … Hanyang University NVRAMOS’11, Jeju, Korea 18/68 Power consumption vs. Channels, Ways, and Clusters 19/68 Read vs. Write =Page Operation =Operation Complete Read C R Write … … … … … … … … C D D P )Am(tnerruC )Am(tnerruC time time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 20/68 Writing 4 pages: 1 Channel X 1 Way =Page Write =Write Complete C D … … … … P )Am(tnerruC time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 21/68 Way switch vs. Channel switch Delay =Page Write =Write Complete Way Switch C D Channel Switch … … … … … … … … C D P )Am(tnerruC )Am(tnerruC Way Switch P Channel Switch time time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 22/68 Writing 4 pages: 1 Channel X 2 Way =Page Write =Write Complete C D … … … … P )Am(tnerruC time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 23/68 Writing 4 pages: 2 Channel X 1 Way =Page Write =Write Complete C D … … … … P )Am(tnerruC time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 24/68 Writing 4 pages: 2 Channel X 2 Way =Page Write =Write Complete C D … … … … P )Am(tnerruC Way Switch Channel Switch time Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 25/68 Cluster p Cluster : Write Unit of SSD =Page Write 4KB Write Youjip Won 8KB Write … … … … Hanyang University = … … … … NVRAMOS’11, Jeju, Korea 26/68 Trivia in Measurement Methodology 27/68 Trivia of Measurement Methodology Sampling interval should be smaller than Read/Program Latency p Smoothing the data to filter out measurement noise. Smoothing Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 28/68 Trivia of Measurement Methodology p Deactivate DRAM cache w SATA command 82h ¢ Samsung MXP and Intel X-25M SSDs w Trigger DRAM flush ¢ Youjip Won OCZ Vertex and Hanamicron Forte+ Hanyang University NVRAMOS’11, Jeju, Korea 29/68 Case Studies 30/68 Measurement Model Vendor Size Channels DRAM Size Package Type X25M Intel 80GB 10 16MB 20 MLC MXP SAMSUNG 128GB 8 128MB 16 MLC Vertex OCZ 60GB 8 64MB 16 MLC Forte+ Hanamicron 32GB 8 32MB 8 MLC p Oscilloscope (TDS3032) p High resolution current probe(TCP202) p Current probe to power line(Vdd) of the SSD p Sampling interval(10samples MA): 10usec p Each request measured 10 times Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 31/68 Case Study : Intel X25M Intel X25M Capacity 80GB No. of Channels 10 Packages/Channel 2 Package 4 GB … … … … … … … … … … … … … … … … … … … … NAND (4GB) Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 32/68 Case Study : Intel X25M Write p IO Size: 4KB to 80KB 600 μsec Increase 20 steps Channel switch = 30 μsec NAND programming: 17 mA Simple Round Robin 17mA 100mA Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 33/68 Case Study : Intel X25M Write p 4KB to 80KB Sequential Write = Page Write Channel Switch =30 μsec 10 11 1213141516171819 0 1 2 3 4 5 6 7 8 9 MUX … … … … … … … … … … … … … … … … … … … … Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 34/68 Case Study : Intel X25M Write p IO Size: 80KB to 160KB Write ? 84 KByte 80 KByte Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 35/68 Case Study : Intel X25M Write =1 Page Write 10 11 1213141516171819 20 0 1 2 3 4 5 6 7 8 9 =Write Complete MUX … … … … … … … … … … … … … … … … … … … … tnerruC w ol F P20 P19 P18 P17 P16 P15 P14 P P13 P P12 P11 P10 P9 P8 P7 P6 P5 P4 P3 P2 P1 P0 time Youjip Won Hanyang University time NVRAMOS’11, Jeju, Korea 36/68 Case Study : Intel X25M Read p IO Size: 4KB to 16KB Read Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 37/68 Case Study : Intel X25M Read p IO Size: 16KB to 28KB Read 180mA 160mA Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 38/68 Case Study : Intel X25M Read p IO Size: 28KB to 48KB Read Same Peak Current Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 39/68 Case Study : Intel X25M Read p IO Size: 48KB to 52KB Read Same Peak Current Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 40/68 Case Study : Intel X25M Read p IO Size: 56KB to 72KB Read Same Peak Current Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 41/68 Case Study : Intel X25M Read p IO Size: 76KB to 92KB Read Same Peak Current Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 42/68 Case Study : Intel X25M Read p IO Size: 96KB to 100KB Read Same Peak Current Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 43/68 Case Study : Intel X25M Summary ü Cluster Size : ü Operation: Write Read 4KB 48KB 17mA ü Peak Current : 500mA ü Channel switch: 30sec ü Policy p p Youjip Won 180mA Round Robin Different from Write, peak current remains the same. Only duration changes. Hanyang University NVRAMOS’11, Jeju, Korea 44 Case Study : SAMSUNG MXP Samsung MXP Capacity 128GB No. of Channels 8 Packages/Channel 2 Package 8 GB … … … … … … … … … … … … … … … … NAND (8GB) Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 45/68 Case Study : SAMSUNG MXP p Write Operation : From 4Kb to 32KB Peak Current 50mA Youjip Won Hanyang University 100mA NVRAMOS’11, Jeju, Korea 46/68 Case Study : SAMSUNG MXP p Write Operation : From 36Kb to 64KB Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 47/68 Case Study : SAMSUNG MXP p Write Operation : From 68Kb to 96KB Same Peak Current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 48/68 Case Study : SAMSUNG MXP p Write Operation : 100Kb to 112KB Same Peak Current Cluster Size of MXP is 32KB Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 49/68 Case Study : SAMSUNG MXP What we still do not know of Peak of 128 Kbyte < Peak Current of 124 KByte Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 50/68 Case Study : OCZ Vertex OCZ Vertex Capacity 60GB No. of Channels 8 Packages/Channel 2 Package 4 GB … … … … … … … … … … … … … … … … NAND (4GB) Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 51/68 Case Study : OCZ Vertex p write-caching off, 1KB write, 10 counts Youjip Won 20ms interval time 100ms interval time 200ms interval time 500ms interval time Hanyang University NVRAMOS’11, Jeju, Korea 52/68 Case Study : OCZ Vertex < 100ms Cache Flush 64MB Write Cache is filled with 64MB write request Test I/O Size Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 53/68 Case Study : OCZ Vertex p IO Size: 4KB to 32KB Peak Current Change 160mA 40mA Same Peak Current 120mA Cluster Size of Vertex is 16KB Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 54/68 Case Study : OCZ Vertex p write-caching off, 1counts, IO Size: 64k to 512k 350mA 220mA Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 55/68 Case Study : OCZ Vertex p IO Size: 4KB to 4096KB Write 350mA 240mA 150mA 120mA 110mA Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 56/68 Case Study : Hanamicron Forte+ Hanamicron Forte+ … Capacity 32GB No. of Channels 8 Packages/Channel 1 Package 4 GB … … … … … … … NAND (4GB) Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 57/68 Case Study : Hanamicron Forte+ p IO Size: 4KB to 32KB Peak Current Change 40mA Same Peak Current Cluster Size of Forte+ is 16KB Size of cluster same with OCZ Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 58/68 Case Study : Hanamicron Forte+ p IO Size: 4KB to 512KB Write 220mA 190mA 160mA 120mA Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 59/68 Comparison: Write Energy p Energy Consumption Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 60/68 Comparison: Peak and Duration of 512K Write 500mA 350mA 220mA Hanamicron Intel OCZ Samsung Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 61/68 Forth Coming Problem in Multi-channel SSD 2008 Read Write 2009 2010 Capacity 2011 Model Name Youjip Won Hanyang University channel 2012~ Release SATA3035 (Mtron) 2008.01 Vertex (OCZ) 2009.03 Vertex2 (OCZ) 2010.07 REVO Drive X2 (OCZ) 2011.01 NVRAMOS’11, Jeju, Korea 62/68 Forth Coming Problem in Multi-channel SSD p 10 channel: peak 500 mA p 16 Channel: peak 800 – 900 mA à SSD is no long Green. p Further, excessive peak current can cause… üsupply voltage drop üground bounce üsignal noise üblack-out üEtc… Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 63/68 Power Budget Power 8 Channels X 1 Way Feasible Solutions Power Budget 2 Channels X 4 Ways Performance Bound 4 Channels X 2 Ways Feasible Region 1 Channel X 8 Ways I/O Latency Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 64/68 Summary X25M MXP OCZ Hanamicron ü Cluster Size : 4KB 32KB 16KB 16KB ü programming: 17 mA 35mA 15~20mA 15~20mA ü Peak Current : 500mA 350mA 350mA 220mA ü Channel switch: 30 μsec ü Energy ü Performance ü Small Write ü Large write ü Standby current Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 65/68 Summary p Energy Consumption is very good tool to characterizeSSD. p For larger number of channels, peak current will soon be a significant issue. p Youjip Won We introduce the notion of Power Budge to resolve this issue. Hanyang University NVRAMOS’11, Jeju, Korea 66/68 Youjip Won Hanyang University NVRAMOS’11, Jeju, Korea 67/68