Intel Optane Data Center Persistent Memory
5 likes3,887 views
The document discusses Intel's Optane DC Persistent Memory architecture and its performance characteristics. It highlights key features like reduced latency, large memory capacity, and the ability to maintain persistency, offering insights into different usage modes and performance benchmarks. Additionally, it emphasizes the importance of system configuration and workload characteristics on performance outcomes.
Intel Optane Data Center Persistent Memory
- 1. Intel®Optane™DataCenter PersistentMemory Architecture (Jane) and Performance (Lily) Presenters: Lily Looi, Jianping Jane Xu Co-Authors: Asher Altman, Mohamed Arafa, Kaushik Balasubramanian, Kai Cheng, Prashant Damle, Sham Datta, Chet Douglas, Kenneth Gibson, Benjamin Graniello, John Grooms, Naga Gurumoorthy, Ivan Cuevas Escareno, Tiffany Kasanicky, Kunal Khochare, Zhiming Li, Sreenivas Mandava, Rick Mangold, Sai Muralidhara, Shamima Najnin, Bill Nale, Jay Pickett, Shekoufeh Qawami, Tuan Quach, Bruce Querbach, Camille Raad, Andy Rudoff, Ryan Saffores, Ian Steiner, Muthukumar Swaminathan, Shachi Thakkar, Vish Viswanathan, Dennis Wu, Cheng Xu 08/19/2019
- 2. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit http://www.intel.com/benchmarks . Configurations on slides 18 and 20. Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service activation. Performance varies depending on system configuration. Intel technologies may require enabled hardware, specific software, or services activation. Check with your system manufacturer or retailer. Performance results are based on testing as of Feb. 22. 2019 and may not reflect all publicly available security updates. See configuration disclosure for details. No product or component can be absolutely secure. Results have been estimated or simulated using internal Intel analysis or architecture simulation or modeling, and provided to you for informational purposes. Any differences in your system hardware, software or configuration may affect your actual performance. Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. For more complete information about performance and benchmark results, visit http://www.intel.com/benchmarks . Cost reduction scenarios described are intended as examples of how a given Intel-based product, in the specified circumstances and configurations, may affect future costs and provide cost savings. Circumstances will vary. Intel does not guarantee any costs or cost reduction. Intel does not control or audit third-party benchmark data or the web sites referenced in this document. You should visit the referenced web site and confirm whether referenced data are accurate. All information provided here is subject to change without notice. Contact your Intel representative to obtain the latest Intel product specifications and roadmaps. *Other names and brands may be claimed as property of others. Intel, the Intel logo, Xeon, the Xeon logo, Optane, and the Optane logo are trademarks of Intel Corporation in the United States and other countries. © Intel Corporation. Notices&Disclaimers
- 3. Intel®Optane™DC PersistentMemory Architecture A Breakthrough with a New Interface Protocol, Memory Controller, Media, and Software Stack 1.
- 4. Memory-StorageGap DRAM HOT TIER HDD / TAPE COLD TIER SSD WARM TIER Intel®3DNandSSD CPU LLC Core L2 L1 pico-secs nano-secs Memory Sub-System 10s GB <100nanosecs Network Storage SSD 10s TB <100millisecs 10s TB <100microsecs Memory-Storage Gap? less often → AccessDistribution Data Access Frequency Cooler data more often Hot data
- 5. CloseMemory–StorageGap less often → AccessDistribution Data Access Frequency Cooler data more often Hot data DRAM HOT TIER HDD / TAPE COLD TIER SSD WARM TIER Optimize performance given cost and power budget CPU LLC Core L2 L1 pico-secs nano-secs Memory Sub-System 10s GB <100nanosecs Network Storage SSD 10s TB <100millisecs 10s TB <100microsecs Move Data Closer to Compute Maintain Persistency 100s GB <1microsec 1s TB <10microsecs Intel®3DNandSSD
- 6. High Resistivity – ‘0’ Low Resistivity – ‘1’ Attributes + Non-volatile + Potentially fast write + High density + Non-destructive fast read + Low voltage + Integrate-able w/ logic + Bit alterable Cross-Point Structure Selectors allow dense packing And individual access to bits Scalable Memory layers can be stacked in a 3D manner Breakthrough Material Advances Compatible switch and memory cell materials High Performance Cell and array architecture that can switch fast First Generation Capacities: 128 GB 256 GB 512 GB Intel®Optane™MediaTechnology
- 7. DRAM 1. DQ buffers presents a single load to the host 2. Host SMBus: SPD visible to the CPU, Optane Controller plays thermal sensing (TSOD) functionality 3. Address Indirection Table 4. Integrated PMIC controlled Optane Controller 5. On DIMM Firmware storage 6. On-DIMM Power Fail Safe with auto-detection Optane Controller NVM PMICPMIC Bus FlashFlush I/F NVM Bus Data Bus Buff Buff Buff Buff Buff Buff SMBus/I2Cs NVMNVMNVMNVMOptane Media CMD and Address Bus SPD NVM Rails Mem Ctrl Rails 1 4 5 2 DDR4Connector DQ Buffer and Logic Rail 6 3 Intel®Optane™DCPersistentMemoryModuleArchitecture
- 8. Interface to Host CPU DCPMMMemoryInterface Addr Mapping Cache Address Mapping Logic Media Management Power & Thermal Mgmt Uctrl Encrypt/ Decrypt Key Mgmt DRNG Scheduler Read Queue Write Queue Refresh Engine ECC/ Scrambler Error Handling Logic OptaneTM MediaInterface Media Channel OptaneTM Media Devices Caps for Flushes Intel®Optane™DCPersistentMemoryControllerArchitecture DDR4Sloton HostCPU
- 9. PersistentMemory USERSPACEKERNELSPACE Standard File API GenericNVDIMMDriver Application FileSystem ApplicationApplication Standard Raw Device Access Load/ Store ManagementLibrary ManagementUI Standard File API pmem-Aware FileSystem MMU Mappings file memory “DAX” Intel®OptaneDCPersistentMemorySWEnablingStack
- 10. CPU Core 1. AC power loss to de-assert the PWROK 2. Platform logic then asserts the ADR_Trigger 3. PCH starts the ADR programmable timer 4. PCH assertion to SYNC message 5. PCU in processor detects SYNC message bit and sends AsyncSR to MC 6. MC flushes Write pending queue (WPQ) 7. After ADR timer expires, PCH asserts ADR_COMPLETE pin Asynchronous DRAM Refresh (ADR) - Power Fail Protected Feature Platform Power Supply Platform Logic PCH ADR_SYNC ADR_GPIO ADR Timer ADR_Complete PCU MC NVMNVMNVMNVMNVMIntel® Optane™ DC PMM ADR_Trigger PWROK1 2 4 3 5 6 L1 L1 L2 L3 WPQ7 KeyFeatureDeepDive-ADR
- 11. • Large Memory Capacity • No software/application changes required • To mimic traditional memory, data is “volatile” • Volatile mode key cleared and regenerated every power cycle • DRAM is ‘near memory’ • Used as a write-back cache • Managed by host memory controller • Within the same host memory controller, not across • Ratio of far/near memory (PMEM/DRAM) can vary • Overall latency • Same as DRAM for cache hit • DC persistent memory + DRAM for cache miss Core L1 L1 L2 L3 Core L1 L1 L2 Core L1 L1 L2 Core L1 L1 L2 NEAR MEMORY FAR MEMORY Memory Controller MOV NEAR MEMORY FAR MEMORY Memory Controller DRAM DRAM MemoryMode
- 12. • PMEM-aware software/application required • Adds a new tier between DRAM and block storage (SSD/HDD) • Industry open standard programming model and Intel PMDK • In-place persistence • No paging, context switching, interrupts, nor kernel code executes • Byte addressable like memory • Load/store access, no page caching • Cache Coherent • Ability to do DMA & RDMA CPUCACHES Minimum required power fail protected domain: Memory subsystem SW makes sure that data is flushed to durability domain using CLFLUSHOPT or CLWB Core L1 L1 L2 L3 MOV Dram MEMORY Persistent MEMORY Memory Controller DRAM WPQ AppDirectMode–PersistentMemory
- 13. PERFORMANCE Intel® Optane™ DC Persistent Memory for larger data, better performance/$, and new paradigms 2.
- 14. Intel®Optane™DC PersistentMemoryLatency 1000x lower latency Note 4K granularity gives about same performance as 256B Read idle latency Smaller granularity (vs. 4K) Intel Optane DC SSD P4800X Intel DC P4610 NVMe SSD Lower is better For more complete information about performance and benchmark results, visit www.intel.com/benchmarks
- 15. Performance can vary based on • 64B random vs. 256B granularity • Read/write mix • Power level (programmable 12-18W, graph is 18W) Intel®Optane™DC PersistentMemoryLatency Ranges from 180ns to 340ns (vs. DRAM ~70ns) Read idle latency Smaller granularity (vs. 4K) Intel Optane DC SSD P4800X Intel DC P4610 NVMe SSD Read/Write (256B) Read (256B) Read (64B) Read/Write (64B) For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
- 16. DRAM Cache MissDRAM Cache Hit • Good locality means near-DRAM performance • Cache hit: latency same as DRAM • Cache miss: latency DRAM + Intel® Optane™ DC persistent memory • Performance varies by workload • Best workloads have the following traits: • Good locality for high DRAM cache hit rate • Low memory bandwidth demand • Other factors: • #reads > #writes • Config vs. Workload size MemoryModeTransactionFlow CPUCACHES Core L1 L1 L2 L3 NEAR MEMORY FAR MEMORY Memory Controller DRAM CPUCACHES Core L1 L1 L2 L3 NEAR MEMORY FAR MEMORY Memory Controller DRAM https://software.intel.com/en-us/articles/prepare-for-the-next-generation-of-memory
- 17. • Synthetic traffic generator represents different types of workloads • Vary size of buffers to emulate more or less locality • Very large data size (much larger than DRAM cache) causes higher miss rate MemoryModePerformancevs.Locality&Load 17 High BW delivered for high demand WL High demand + poor locality = degradation Medium/low demand WLs still meet requirement 0.0 20.0 40.0 60.0 80.0 100.0 120.0 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% Bandwidth(GB/s) DRAM cache miss rate MLC Bandwidth Varying Bandwidth Demand (100% reads) Light Demand (13GB/s) Med (33GB/s) Heavy Demand (max) For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
- 18. 2-2-2 System Platform Neon city CPU CLX-B0 CPU per Node 28core/socket, 1 socket, 2 threads per core Memory 6x 16GB DDR + 6x 128GB AEP QS SUT OS Fedora 4.20.6-200.fc29.x86_64 BKC WW08 BIOS PLYXCRB1.86B.0576.D20.1902150028 (mbf50656_0400001c) FW 01.00.00.5355 Security Variants 1,2, & 3 Patched Test Date 4/5/2019 MemoryModePerformance/Load/Locality MLC parameters: --loaded_latency –d<varies> -t200 Buffer size (GB) per thread 2-2-2 Miss rate (%) R W2 ~0 0.1 0.1 ~10 1.0 0.7 ~25 4.5 1.8 ~40 9.0 4.5
- 19. 19 1. One Redis Memtier instance per VM 2. Max throughput scenario, will scale better at lower operating point 2 VMs per core2 1 VM per core VM size DRAM baselineMM capacity Throughput vs. DRAM Summary VM’s 45GB 768GB 1TB 111%, meets SLA 42% more VM’s @lower cost 14->20 90GB 768GB 1TB 147%, meets SLA 42% more VM’s@lower cost 7->10 EnableMoreRedisVMInstanceswithSub-msSLA For more complete information about performance and benchmark results, visit www.intel.com/benchmarks. Throughput: Higher is better, Latency: lower is better (must be 1ms or less)
- 20. Configuration 1 - 1LM Configuration 2 – Memory Mode (2LM) Test by Intel Intel Test date 02/22/2019 02/22/2019 Platform Neoncity Neoncity # Nodes 1 1 # Sockets 2 2 CPU Intel® Xeon® Platinum 8276, 165W Intel® Xeon® Platinum 8276, 165W Cores/socket, Threads/socket 28/56 28/56 HT On On BIOS version PLYXCRB1.86B.0573.D10.190130045 3 PLYXCRB1.86B.0573.D10.1901300453 BKC version – E.g. ww47 WW06 WW06 AEP FW version – E.g. 5336 5346 (QS AEP) 5346 (QS AEP) System DDR Mem Config: slots / cap / run-speed 12 slots / 32GB / 2666 12 slots / 16 GB / 2666 System DCPMM Config: slots / cap / run-speed 12 slots / 32GB / 2666 12 slots /128GB,256GB,512GB/ 2666 Total Memory/Node (DDR, DCPMM) 768, 0 192, 1TB,1.5TB,3TB, 6TB NICs 2x40GB 2x40GB OS Fedora-27 Fedora-27 Kernel 4.20.4-200.fc29.x86_64 4.20.4-200.fc29.x86_64 AEP mode: ex. MM or AD-volatile (replace DDR) or AD- persistent (replace NVME) 1LM Memory Mode (2LM) Workload & version Redis 4.0.11 Redis 4.0.11 Other SW (Frameworks, Topologies…) memtier_benchmark-1.2.12 (80/20 read/write) ; 1K record size memtier_benchmark-1.2.12 (80/20 read/write) ; 1K record size VMs (Type, vcpu/VM, VM OS) KVM, 1/VM, centos-7.0 KVM, 1/VM, centos-7.0 RedisConfiguration
- 21. App Direct ReadTraditional Read to Page Fault • Traditional read to page fault (disk): Software 4K transfer from disk Request returned • App Direct access memory directly • Avoids software and 4K transfer overhead • Cores can still access DRAM normally, even on same channel AppDirectModeTransactionFlow CPUCACHES Core L1 L1 L2 L3 NEAR MEMORY Memory Controller DRAM CPUCACHES Core L1 L1 L2 L3 NEAR MEMORY FAR MEMORY Memory Controller DRAM SSD One cacheline 1 2 1 2 3 3
- 22. • Reduce TCO by moving large portion of data from DRAM to Intel® Optane™ DC persistent memory • Optimize performance by using the values stored in persistent memory instead of creating a separate copy of the log in SSD (only pointer written to log) • Direct access vs. disk protocol RedisExample(withSoftwareChange) 22 AOF file (log) AOF Pointers (to log) Write request Store key/ value Append operation into AOF log file Write request Store key Store value Append pointer into AOF log file PMEM Moving Value to App Direct reduces DRAM and optimizes logging by 2.27x (Open Source Redis Set, AOF=always update, 1K datasize, 28 instances) For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
- 23. Configuration Query time 768GB DRAM 1417s 192GB DRAM 1TB App Direct 171s 3TB scale factor 8X improvement in Apache Spark* sql IO intensive queries for Analytics SparkSQLOAPcache • Intel® Optane™ DC persistent memory as cache • More affordable than similar capacity DRAM • Significantly lower overhead for I/O intensive workloads For more complete information about performance and benchmark results, visit www.intel.com/benchmarks.
- 24. • Intel® Optane™ DC Persistent Memory closes the DDR memory and storage gap • Architected for persistence • Provided large capacity scales workloads to new heights • Offered a new way to manage data flows with unprecedented integration into system and platform • Optimized for performance and orders of magnitude faster than NAND • Memory mode for large affordable volatile memory • App Direct mode for persistent memory Summary