Are your ready for in memory applications?
- 1. The Magic and the Mystery of In-Memory Apps Taufik Ma – Industry Insight Shaun Walsh - Marketeer
- 2. 2 © 2015 G2M COMMUNICATIONS. All rights reserved. Contents The Use In Memory Applications? Evolution towards & Role of In-Memory Computing Role of Storage in In-memory solutions Customer Trends Emerging Technologies & Some Predictions Summary
- 3. Magic and In-Memory Applications Shaun Walsh - Marketeer
- 5. 5 © 2015 G2M COMMUNICATIONS. All rights reserved. The Evolution of Storage Tiers NVM will Accelerate Both Meta-Data & Application Data
- 6. 6 © 2015 G2M COMMUNICATIONS. All rights reserved. NVDIMM Acceleration Segments Latency Meta Data Acceleration NVDIMM Type Presentation Access Method Latency -N DRAM Byte Consistent -F Storage Block Variable -P DRAM and/or Storage Byte & Block Variable Meta Data Acceleration NVDIMM-N NVDIMM-F NVDIMM-P 3D-XPoint • Data Base Log Files • Clustering • Cache Synchronization • In-Memory DBs • MemCacheD • RAID • De-Dupe
- 7. 7 © 2015 G2M COMMUNICATIONS. All rights reserved. NVM-DIMM – fills growing DRAM-NAND gap In Memory Applications are driving a new class of Storage Class Memory (SMC) Latency and persistence are as important as absolute bandwidth Byte and Block address flexibility is vital to scaling In-Memory Applications (IMA) 7 Persistent Memory
- 8. 8 © 2015 G2M COMMUNICATIONS. All rights reserved. The Future of Business Intelligence Latency and Persistence are the new value currency for real-time applications & storage • Old performance was data rates (GB/s) & capacity (TB) • Store Everything, Sort Later • Higher Cost, Slow Decisions Latency & PersistenceBandwidth & Capacity • Real-Time is Business Critical • Major Players Driving NMV • Store the Vital & Analyze now Latency & Persistence
- 9. 9 © 2015 G2M COMMUNICATIONS. All rights reserved. Procter & Gamble - Real-Time Reporting & Business Decisions https://hana.sap.com/abouthana/customer-stories/pg.html 35,000 Retail, supply chain and business users supported 400% Increase in decision support systems performance 55% Reduced database from 36TB to 16TB all in memory P&G achieved faster, more reliable reporting and analytics
- 10. McLaren Group – Faster Formula 1 • Faster and more consistent lap times • Improved down force for better grip • Real-time telemetric analysis • More World Championships
- 11. The Art and Science of In Memory Applications Taufik Ma Industry Insight
- 12. 12 © 2015 G2M COMMUNICATIONS. All rights reserved. Evolution of Databases & Analytics 1980s 1990s 2000s-2015 RDBMS EDW/OLAP RDBMS Operational (OLTP, ERM) Data Warehousing (Data mining, DSS, Analytics) RDBMS NoSQL Hadoop EDW/OLAP Oracle, MS SQL, Sybase Teradata, Oracle, SAS, etc MongoDB Cassandra MapReduce HBase MySQL Postgres IBM Netezza EMC Greenplum
- 13. 13 © 2015 G2M COMMUNICATIONS. All rights reserved. RDBMS NoSQL EDW/OLAP Hadoop Structured Data, Relational Unstructured, Schema-less Real-time, Online Operations Batch, Offline Analytics Ongoing Evolution & Specialization… OLTP, ERM Purchases, clicks User profiles, reviews Content Management User Segmentation Daily offer recommendation Ad serving engine Fraud Detection
- 14. 14 © 2015 G2M COMMUNICATIONS. All rights reserved. RDBMS NoSQL EDW/OLAP Hadoop Structured Data, Relational Unstructured, Schema-less Real-time, Online Operations Batch, Offline Analytics Ongoing Evolution & Specialization… In-Memory Database Hana, Exalytics, MemSQL, etc In-Mem Data Processing Spark, Hadoop in-mem Real-time analytics OLTP, ERM Purchases, clicks User profiles, reviews Content Management Financial risk/value analysis Fraud Prevention Real-time recommendations Profitability analysis User Segmentation Daily offer recommendation Ad serving engine Fraud Detection
- 15. 15 © 2015 G2M COMMUNICATIONS. All rights reserved. Multiple Tools Within A Customer Customer Profiles (G2M Survey) $500M+ Retail $500M+ Pharma $1B+ Manufacturing $1B+ Pharma $1B+ SaaS $250M+ Healthcare Hadoop Yes Yes Yes Yes Yes Yes MongoDB Yes No plans Yes Yes No plans Spark Yes No plans Considering Yes, in 6 months Yes Yes, in 6 months SAP HANA No plans Yes Considering Yes No plans Considering Microsoft Hekaton No plans No plans Considering Yes, in 6 months No plans Yes, in 12 months memSQL No plans No plans Considering Yes, in 6 months No plans Yes, in 12+ months Oracle Exalytics No plans No plans Yes Yes No plans Yes, in 12+ months “Specialized Tools for Specific Needs” (Or “Too Many Data Islands”?)
- 16. 16 © 2015 G2M COMMUNICATIONS. All rights reserved. Multiple In-Memory Applications within a Customer How many in-memory applications do you (or will you) run? 1-5 6-10 More than 10
- 17. 17 © 2015 G2M COMMUNICATIONS. All rights reserved. Key Enabler of In-Memory Computing: Today’s Technologies On a human scale… If I complete 50 operations in 50 seconds, then have to wait for data… Time to get data CPU L1 cache 0.001 usec DRAM 0.01 usec NAND 100 usec HDD 10,000 usec DRAM = getting food from the fridge (10’s of seconds) NAND = taking the day off HDDs = hiking the Pacific Coast Trail (months)
- 18. 18 © 2015 G2M COMMUNICATIONS. All rights reserved. Performance Comes at a Price Storage Time to get data Price / GB Cost for 100TB # 2U Servers Req’d to Hold 100TB* DRAM 0.01 usec $5.60 32G DIMM for $179 ea, Samsung Registered DDR4, M393A4K40BB0-CPB0 $560,000 3125 x 32G DIMMs 130 NAND 100 usec $0.35 2.5” 1TB SSD, $350 ea, Intel 540S $35,000 100 x 2.5” 1TB SSD 5 HDD 10,000 usec $0.03 3.5” 4TB SATA HDD for $120 ea, Seagate ST4000DM000 $3,000 25 x 3.5” 4TB SATA HDD 2-3 * Assuming 24 DIMM slots, 24x 2.5” drives or 12x 3.5” drives
- 19. 19 © 2015 G2M COMMUNICATIONS. All rights reserved. Location of Data & Tasks Input File Chunks 1 2 3 Hadoop: MapReduce / HDFS Parallel Tasks DISK 1 Parallel Tasks Parallel Tasks 2 3 Input File Partitions (RDDs) 1 2 3 Spark / Tachyon 1 Parallel Tasks MEM 2 Parallel Tasks 3 Parallel Tasks Input File User Partitioning 1 2 SAP Hana 1 Local Tasks MEM 2 Local Tasks Master Slave(s) StandbyJobTracker / Name Node Sends tasks to data nodes Spark Driver Sends tasks to worker nodes
- 20. 20 © 2015 G2M COMMUNICATIONS. All rights reserved. Surviving Failures Input Files Chunks 1 2 3 Hadoop: MapReduce / HDFS Parallel Tasks DISK 1 3 Parallel Tasks Parallel Tasks 2 3 3 2 Input Files To persistent storage 1 2 3 Spark / Tachyon 1 Parallel Tasks MEM 2 Parallel Tasks 3 Parallel Tasks Input Files User Partitioning 1 2 SAP Hana 1 Local Tasks MEM 2 Local Tasks Ext Storage Logs & savepoints Lineage: Record of transformations that created an RDD from its “parent” 3-fold Replication 2 1 1 Lineage Master Slave(s) Standby & checkpoints Partitions (RDDs)
- 21. 21 © 2015 G2M COMMUNICATIONS. All rights reserved. No such thing as 100% In-Memory a b c Input Files Chunks 1 2 3 Hadoop: MapReduce / HDFS 1 Parallel Tasks DISK* 2 3 SSD RAM_ DISK a 2 Parallel Tasks a 3 Parallel Tasks a 1 3 2 1 a b c Input Files Partitions (RDDs) 1 2 3 Spark / Tachyon 1 Parallel Tasks HDD SSD MEM a 2 Parallel Tasks a 3 Parallel Tasks a HDFS2.0 Heterogeneous Storage Storage Types & Policies Files/directories assigned policies (e.g. Lazy_persist, All_SSD) Tachyon Tiered Storage (for Off_heap Spark RDDs) Auto or manual a b c * ARCHIVE tier not shown a b cb a ac c b a b Input Files User Partitioning 1 2 SAP Hana 1 Local Tasks MEM 2 Local Tasks Ext Storage a b Logs & savepoints Caching WARM: Primary image on Disk HOT: Primary image in Mem SAP HANA Dynamic Tiering Data spec’d as either Hot or Warm
- 22. 22 © 2015 G2M COMMUNICATIONS. All rights reserved. Customer In-Memory Computing Trends (based on G2M survey) • Cluster sizes similar to big data solutions o ½ respondents > 500 servers, 1/3 at >50 o And not just for Spark • With datasets that fit available DRAM capacity o 1/3 at >100TB, 1/3 at >10TB ~Half with 10-20%+/yr dataset growth Majority use/want tier-ing when dataset > DRAM Only minority would rely on scale- out only Mixed on whether tier-ing should be transparent or not Some want it transparent to developer; Rest want developer to have control via policy • ~Half believe “my storage capacity forces me to have more compute capacity then I need” • Majority have or have plans for consolidated data silos o OLTP+IMDB, Spark+Hadoop, NoSQL+Hadoop SIZE GROWTH EFFICIENCY
- 23. 23 © 2015 G2M COMMUNICATIONS. All rights reserved. Emerging Technologies: High-speed Fabrics & Disaggregated Storage Ethernet or PCIe based fabric DAS-like performance Local or SAN Map any drive to any host Scale each storage tier separately from compute Early proof points: EMC DSSD, SanDisk InfiniFlash, DriveScaleHDD NAND DRAM Low latency fabric CPU CPU CPU CPU CPU CPU 50G 40G 25G 10G 100G … Data Center Ethernet speeds ramping faster than drive speeds: 10/25/40/50/100G RDMA-over-Ethernet technologies Multi-host PCIe fabrics emerging (e.g. OCP Lightning) albeit w/ less scalability SATA/SAS NVMe PCIeX4 Gen3 time
- 24. 24 © 2015 G2M COMMUNICATIONS. All rights reserved. Emerging Technologies: Storage Class Memory Storage Persist- ence Time to access data Price / GB Cost for 100TB # 2U Servers Req’d to Hold 100TB* DRAM N 10ns+ $5.60 $560,000 3125 x 32G DIMMs 130 NV-DIMM -N Y 10ns+ $10+ If 2X+ DRAM $1,000,000+ 260 16G NVDIMM, supercap 3DXP DIMM 100ns Rd 500ns Wr $2+ If 1/3+ DRAM $190,000+ ~50 assuming 96 or 128GB DIMMs NAND Y 100 usec $0.35 2.5” 1TB SSD, $350 ea, Intel 540S $35,000 100 x 2.5” 1TB SSD 5 HDD Y 10,000 usec $0.03 3.5” 4TB SATA HDD for $120 ea, Seagate ST4000DM000 $3,000 25 x 3.5” 4TB SATA HDD 2-3 * Assuming 24 DIMM slots, 24x 2.5” drives or 12x 3.5” drives
- 25. 25 © 2015 G2M COMMUNICATIONS. All rights reserved. In-Memory Computing Predictions / Trends 1. 3DXP DIMMs used for “Jumbo Memory” – value in lower $/GB vs DRAM, not persistence – Mix of 3DXP & DRAM DIMMs in server nodes – Tier-ing will be tuned to accommodate slower writes & reads – Spark, In-mem Hadoop, MemSQL, Hana, etc – NV-DIMM –P might have similar adoption but predictable latency is a concern 2. Increasing use of NVMe SSDs as “Far Memory” – as next tier (below DRAM/3DXP) – Priority on $/TB, not persistence. Resiliency still via Lineage, logs, etc – Remove ”last-inch” of latency via BLKB (block-layer/kernel bypass) stacks (e.g. EMC libflood, SPDK) – Implemented as a fabric-disaggregated cluster to enable efficiency & independent scalability – Longer-term, HW-based paging of near-memory to far-memory 3. Use of “Persistent Memory” for In-Mem computing will evolve – For 3DXP & NV-DIMM –N – Industry progress on pmem file systems (Linux, Windows) – Does persistence replace or complement lineage/logs? – Need low latency replication across nodes (PMoF)
- 26. 26 © 2015 G2M COMMUNICATIONS. All rights reserved. Summary In-memory solutions growing in adoption – driven by real-time analytics Co-existence of structured (e.g. Hana) and unstructured frameworks (e.g. Spark) Confluence of big-data & real-time analytics drives increasing adoption of tier-ing Newer technologies on horizon will continue to create disruptions to in-memory computing architectures