How to Get a Game Changing Performance Advantage with Intel SSDs and Aerospike
Download as PPTX, PDF1 like1,497 views
The document discusses various non-volatile memory solutions and their applications, highlighting key players and technologies in the database realm, including Aerospike and Intel SSDs. It provides performance metrics for different configurations of Aerospike databases and outlines areas for improvement in existing systems, focusing on I/O wait times and software interrupts. Additionally, it mentions ongoing innovation projects between Intel and Aerospike aimed at optimizing database performance.
How to Get a Game Changing Performance Advantage with Intel SSDs and Aerospike
- 1. Frank Ober Solutions ArchitectureTeam – Non-Volatile Memory Solutions Group @fxober
- 6. Type Speed Usage Players Key value databases Fastest Operational MemcacheDB, Redis, Aerospike DynamoDB(Amazon). LevelDB (Google), Rocksdb (Facebook) BigTable , Column- based. Faster Analytics BigTable, Cassandra, Hbase (*Hadoop) Document Databases Faster Web Sites MongoDB (WiredTiger v3.0 is released) Couchbase (ForestDB releases June 2015) Graph databases Fast Social Graphs Neo4J
- 7. Level DB and how SSTable works - https://www.igvita.com/2012/02/06/sstable-and-log-structured-storage-leveldb/ Flash and Storage Filesystems - http://radar.oreilly.com/2014/08/how-flash-changes-the-design-of-database-storage- engines.html (sponsored by Aerospike) NVM KV – Store (Read Side is Direct): NVM KV Store is asynchronously written through a memory buffer. AerospikeConf file setting: write_block_size 128k is default Post-write-queue 256 is default
- 9. 3 Clients You need to spread the load Here Dell 620 dual sockets are used DUAL 10Gbit networks Dell R730xd Server System One primary (dual system with replication testing) DualCPU socket, rack mountable server system DellA03 Board, Product Name: 0599V5 CPU Model used 2 each - Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz max frequency: 4Ghz 18 cores, 36 logical processors per CPU 36 cores, 72 logical processors total DDR4 DRAM Memory 128GB installed BIOSVersion Dell* 1.0.4 , 8/28/2014 Network Adapters Intel® EthernetConverged 10G X520 – DA2 (dual port PCIe add-in card) 1 – embedded 1G network adapter for management 2 – 10GB port for workload Storage Adapters None Internal Drives and Volumes / (root) OS system – Intel SSD for DataCenter Family S3500 – 480GB Capacity /dev/nvme0n1 Intel SSD for Data Center Family P3700 – 1.6TBCapacity, x4 PCIe AIC /dev/nvme1n1 Intel SSD for DataCenter Family P3700 - 1.6TBCapacity, x4 PCIe AIC /dev/nvme2n1 Intel SSD for Data Center Family P3700 - 1.6TB Capacity, x4 PCIe AIC /dev/nvme3n1 Intel SSD for DataCenter Family P3700 - 1.6TBCapacity, x4 PCIe AIC 6.4TB of raw capacity for Aerospike database namespaces Aerospike Community Version 3.5.8
- 10. https://communities.intel.com/community/itpeernetwork/blog/2015/02/17/reaching-one- million-database-transactions-per-second-aerospike-intel-ssd Record Size Aerospike Number of clients threads Total TPS Percent below 1ms (Reads) Percent below 1ms (Writes) Std Dev of Read Latency (ms) Std Dev of Write Latency (ms) Approx. Database size 1k 576 1,124,875 97.16 99.9 0.79 0.35 100G 2k 448 875,446 97.33 99.57 0.63 0.18 200G 4k 384 581,272 97.22 99.85 0.63 0.05 400G 1k (replication) 512 1,003,471 96.11 98.98 0.87 0.30 200G Record Size iostat Read MB/sec Write MB/sec Avg queue size on SSD Average drive latency CPU Busy % 1k 418 29 31 0.11 93 2k 547 43 27 0.13 81 4k 653 52 20 0.16 52 1k (replication) 396 51 30 0.13 94
- 11. StorageTypes Cost per GB 1k transaction/socket Memory Capacity DRAM only $10-15 + (DDR4) ~1.6 million tps (1 socket) 192GB – 768 GB SSD Configuration $1-3 + (PCIe SSD – retail channel) ~600k per node (1 socket) 4 x 2TB = 8TB 10# SFF NVMe servers 3x lower transactions per second, yet 5x lower price per GB with NVM. Capacity is higher, cost is much lower allowing you to do more per unit of rack.
- 12. Areas of improvement exist: 1. Iowait percentage is pretty high here: 13%, and because all SSD’s degrade with writes, you really want to achieve the most that you can with good SSD’s. Endurance is not the issue here, what is the issue is the NAND characteristics related to writes and overall (R+W) latency with write characteristics. 2. Software IRQ (14%)– this is an opportunity to help balance the network and storage drivers more effectively over time. 1. There are several innovation projects going on between Intel – Aerospike related to optimizations…
- 14. asadm (asmonitor) top, htop (mpstat) iostat iotop –o iftop, tcpdump/Wireshark http://www.tecmint.com/command-line-tools-to-monitor-linux-performance/ Intel CPU monitor (PCM) vtune mlc-