IMCSummit 2015 - Day 2 IT Business Track - Drive IMC Efficiency with Flash Extended Memory
0 likes695 views
The document discusses advancements in driving computing efficiency with flash extended memory and the implications for applications. It highlights the performance improvements of non-volatile memory (NVM) technologies, specifically in optimizing database performance for workloads through the integration of flash technology. The presentation emphasizes the need for applications to adapt to flash architecture to enhance performance and lower costs, detailing use cases and benefits of 'flash-aware' systems.
IMCSummit 2015 - Day 2 IT Business Track - Drive IMC Efficiency with Flash Extended Memory
- 1. 1 Driving IMC Compu/ng Efficiency with Flash Extended Memory Dan Baigent, Sr Director Strategic Partner Ecosystems June 30, 2015
- 2. 2 Forward-‐Looking Statements During our mee;ng today we will make forward-‐looking statements. Any statement that refers to expecta;ons, projec;ons or other characteriza;ons of future events or circumstances is a forward-‐looking statement, including those rela;ng to market growth, products and their capabili;es, performance and compa;bility, cost savings and other benefits to customers. Informa;on in this presenta;on may also include or be based upon informa;on from third par;es, which reflects their expecta;ons and projec;ons as of the date of issuance. We undertake no obliga;on to update these forward-‐looking statements, which speak only as of the date hereof.
- 3. 3 Widening Performance Gap 1990 1995 2000 2005 2010 2015 Server CPU Performance Mega Trend: Legacy Storage I/O BoKleneck Aligning Performance SSD Performance HDD Performance Source: StorageIOblog; Sep 2009; hSp://storageioblog.com/data-‐center-‐io-‐boSlenecks-‐performance-‐issues-‐and-‐impacts/
- 4. 4 Non-‐Vola/le Memory (NVM) ! Today: NAND Flash – Capacity: 100s of GB to 100s of TB per device – Trends: Higher capacity, lower cost/GB, lower write cycles, SLC-‐>MLC-‐>3BPC – IOPS: 100K to millions, GB/s of bandwidth ! Tomorrow: Non-‐Vola;le Memory technologies (Phase Change Memory, MRAM, STT-‐RAM, etc.) – Poten;al for orders of magnitude performance improvement Fusion ioMemory™ PCIe SAS and SATA SSDs InfiniFlash™ System SanDisk ION Accelerator™
- 5. 5 Why Do Applica/ons Need Op/miza/on for Flash? ! Many applica;ons assume high latency storage (some even op;mize for read/ write head posi;oning) ! Flash is different from disk – Performance, endurance, addressing – Geeng more different over ;me ! Flash-‐focused applica;on accelera;on – Shifing boSlenecks (Compute to I/O to Network to Applica;on) – Managed writes = greater device life;me (wear leveling, endurance) – Improved system efficiency (TCO and TCA) – Even lower power and cooling costs Area Hard Disk Drives Flash Devices Read/Write Performance Largely symmetrical Heavily asymmetrical. Sequen;al vs Random Performance 100x difference. <10x difference. Background ops Rare Regular Wear out Largely unlimited Limited writes IOPS 100s to 1,000s 100Ks to Millions Latency 10s milli sec 10s-‐100s micro sec Addressing Sequence, Sector Direct, byte addressable
- 6. 6 Becoming “Flash Aware”: SanDisk NVMFS Value ! Increase life expectancy of flash devices ! Consistent low latency ! Consistent high performance How ! Reducing MySQL™ Writes to flash ! Op;mize IO Write path for flash ! Applica;ons leverage enhanced I/O interface Available today! Non-‐Vola;le Memory File System – Op;mized for Flash and Beyond Native Flash Translation Layer block allocation, mapping, recycling ACID updates, logging/journaling, crash-recovery NVMFS file metadata mgmt Kernel block layer kernel-space user-space Standard Linux File Systems file metadata mgmt, block allocation, mapping, recycling, ACID updates, logging/journaling, crash-recovery Flash Primitives MySQL™ Database Linux VFS (virtual file system) abstraction layer New File system Elimina;ng Duplicate Logic and Leveraging New Primi;ves for Op;mal Flash Performance and Efficiency
- 7. 7 Persistent Storage In-‐Memory Database and Persistence Checkpoints, Logs and Data Tiering, Oh My! Logs Mem Image Checkpoints State Changes Transac'ons over Time Warm Data Data Tiering Backup Systems In-‐Memory Compute (transac/ons, processing, analy/cs) “Hot Data” Persistent Snapshots Transac;on Commit Restore Data into memory Backup and Archiving Possible wait condi;ons Logs
- 8. 8 Emergence of Flash Extended Memory Flash DevicesMemory Persistent Memory HDDs Memory Addressable and I/O Addressable “Flash as Memory” “Flash as Disk” Tradi;onal Block I/O Storage Device Technologies: Flash Extended Memory Decoupling of the physical infrastructure (memory) from the management and u/liza/on of that infrastructure through a so[ware interface
- 9. 9 Flash Extended Memory Example: MongoDB – Lowering TCO through transparent DRAM/Flash /ering MongoDB Throughput (opera;ons/second) 0 2000 4000 6000 8000 24 GB Node 12 GB Node 8 GB Node -‐26% -‐33% Throughput Devia;on (opera;ons/second) 0 100 200 300 400 24 GB Node 12 GB Node 8 GB Node 3x Improvement YCSB Workload A DRAM reduc;ons of 2x and 3x yield 26% and 33% throughput degrada;ons respec;vely. Throughput predictability actually improves with less DRAM. Latency shows similar trend. Source: Based on internal tes;ng by SanDisk; Nov 2014
- 10. 10 RA New Memory-‐Storage Hierarchy Range of persistence – I/O becomes Load-‐ Store Storage Memory Convergence: Rethinking the Memory Hierarchy L1, L2, L3 CPU Caches DRAM Persistent Memories Flash Hard Drive Microseconds Nanoseconds CYCLES TO WAIT Main Memory System High Performance I/O System Accessed Like Memory and Managed Like Storage Milliseconds ACCESS DELAY 2 cycles 4 million cycles
- 11. 11 Implica/ons for Applica/ons ! Extremely low latency transac;on commits – 10-‐100us reduced to <1us ! Accelerate logs, indexes etc. ! Convergence of in-‐memory and disk approaches – New data structures to op;mize directly for memory access ! Rich variety of programming models – Tradi;onal I/O will s;ll be available – Mmap directly to persistent memory is alterna;ve model: use pointer opera;ons to manipulate persistent data – SNIA Non-‐Vola;le Memory Programming TWG – standardize mmap for persistent memory ! New challenges – CPU cache management – Atomicity?
- 12. 12 Technology Preview: Database Accelera/on Through Flash Extended Memory NVMFS (POSIX compliant file system) ACM (Auto-‐Commit Memory) So[ware • A “transparent” Sofware Defined Memory layer can provide accelerated I/O over a “baseline” unaware interface • But “flash-‐aware” applica;ons can op;mize that accelera;on Flash DevicesMemory Persistent Memory Memory Device Technologies: Memory Interfaces Oracle MySQL™ (No applica;on changes) (Applica;on op;mized) “Flash -‐Aware” “Transparent” Flash Extended Memory
- 13. 13 Technology Preview Example Configura/on HP ProLiant DL380 Gen8 Fusion ioMemory™ PCIe-‐based flash Standard MySQL database “Baseline” Flash Extended Memory Enabled Standard Flash Accelera/on HP ProLiant DL380 Gen8 Fusion ioMemory™ and persistent memory with NVMFS Standard MySQL database “Transparent” NVMFS and ACM HP ProLiant DL380 Gen8 Op/mized MySQL database “Flash Aware” NVMFS and ACM Fusion ioMemory™ and persistent memory with NVMFS
- 14. 14 Performance Results Latency (lower is beSer) Throughput (higher is beSer) Source: Based on internal tes;ng by SanDisk; Nov 2014
- 15. 15 Advantages & Benefits ! Improve “Baseline” MySQL throughput performance by roughly 60% via “Transparent” accelera;on (no sofware mods) ! Op;mize MySQL throughput performance by over 3x with “Flash Aware” accelera;on (modified sofware) ! Improve “Baseline” MySQL latency by roughly 2.3x (Transparent) and op;mized latency by more than 4x (Flash Aware) ! Uses “Flash-‐as-‐Memory” byte-‐addressable architecture and interface ! Seamless deployment – add ioMemory and NVMFS/ACM sofware to Linux ! Increase performance and capacity in flexible configura;ons Source: Based on internal tes;ng by SanDisk; Nov 2014
- 16. 16 Restore Data into memory Backup and Archiving In-‐Memory Compu/ng – New Approach to I/O Backup Systems In-‐Memory Compute (transac/ons, processing, analy/cs) “Hot Data” Transac'ons over Time Persistence via CPU Load/ Store POSIX File System Persistent Storage Flash Extended Memory Persistent MemoryMemory Flash Devices
- 17. 17 Thank You! @BigDataFlash #bigdataflash ITblog.sandisk.com hSp://bigdataflash.sandisk.com ©2015 SanDisk Corpora;on. All rights reserved. SanDisk is a trademark of SanDisk Corpora;on, registered in the United States and other countries. InfiniFlash, SanDisk ION Accelerator, Fusion ioMemory, ioDrive, Cloudspeed MAX, Op;mus MAX and Lightning Ultra are trademarks of SanDisk Enterprise IP LLC. All other product and company names are used for iden;fica;on purposes and may be trademarks of their respec;ve holder(s).