Pliops - IT Press Tour Israel #43 April 2022

Transcript

1. Scaling in Constrained Environments Multiply Infrastructure Scalability and Efficiency with Pliops XDP Aryeh Mergi, Chairman & Founder Steve Fingerhut President & Chief Business Officer Uri Beitler CEO & Founder IT Press Tour – April 3, 2022 Moshe Twitto CTO & Founder
2. Pliops Overview Customer and Industry Challenges Pliops XDP Explained Customer Use Cases Customer and Partner Endorsements Building More Sustainable Data Centers Topics
3. 3 MISSION To become the NVIDIA of storage by massively increasing infrastructure scaling for databases, analytics, AI/ML, edge, distributed file systems and software- defined storage, and other data-intensive applications HIGHLIGHTED CUSTOMERS SaaS Top 5 SaaS Provider: Core database reliability and user scaling IaaS Top 5 Hyperscaler: Deployed QLC for density and improved CSAT for high- performance elastic block storage HPC Top 500 HPC: File systems and native Key-Value application acceleration INVESTORS
4. The Data-Compute Bottleneck 2010 2020 Performance This Doesn’t Scale! NVMe SSDs CPU 4
5. 5 Data Processing is #1 Focus for Innovation in Data Centers
6. 6 High Demand–Chip Shortage Persists Lead Times Remain Extended 26.2 Weeks1 1 Source: Susquehanna Financial Group https://www.bloomberg.com/news/articles/2022-03-11/wait-times-for-chip-deliveries-grow-again-as-shortages-persist $113.1B $153.9B $190.4B $0 $20 $40 $60 $80 $100 $120 $140 $160 $180 $200 2020 2021 2022 (F) Dollars (B) +68% Semiconductor CapEx Spending2 2 Source: IC Insights: https://www.icinsights.com/news/bulletins/Semi-Industry-Capex-Forecast-To-Jump-24-And- Reach-Over-190-Billion-This-Year/ up 200% from pre-covid Business growth limited by supply, companies making difficult DC infrastructure tradeoffs Industry is trying to catch up – will take years
7. Expanding Demands and Constraints 7 User Growth New Services Data Growth Power & Cooling DC / Rackspace Budget Environmental Responsibility Supply Chain Business Demands Cloud & Enterprise Data Centers Data Center Constraints Current solutions don’t adequately address Supply constraints make it worse Need to rethink data center architecture
8. Pliops Extreme Data Processor (XDP) Performance 3-15x ↑ Efficiency Up to 86% lower TCO Capacity Up to 6x more 8 Reliability DFP* 2x > RAID 0 *Full Performance Drive Fail Protection
9. Industry Landscape – Superior Data Processor Performance/Cost Applicability High Low High Low CPU DPU GPU Computational Storage Pliops = Accelerate All Flash Workloads Built in Data Protection Built in Enhanced Storage Endurance Built in Block & KV Interface Option (No Software Changes)
10. Breakthrough Value vs. Software Performance benefit over software x 1x 2x 3x 4x 5x 6x 7x 8x 9x 10x MongoDB Ceph-QLC MySQL / MariaDB DFP/RAID 5-6 Redis Increased Application Scaling 10 Significant TCO/IOPS Savings TCO Savings over software only -10% 15% 40% 65% 90% Ceph-QLC MongoDB MySQL / MariaDB DFP/RAID 5-6 Redis
11. Broad Use Cases & Market Applicability Target Applications Analytics NoSQL SQL/NewSQL Cloud Instances Block Storage Object DFS Market Segments SaaS Social Media HPC Edge Fintech Cloud IaaS
12. Pliops Innovations Vectors Application Awareness Virtualization Storage Engine Data Services TCO Application Flexibility Data Analytics TP4091 based acceleration Bring application intelligence closer to data Purpose built storage engine with broad vertical integration and benchmarks Data Acceleration Data Reliability Data Compression Data Serialization Data Security Built in multi-DB, multi-volume, SRIOV/SIOV-based virtualization and data services Native Key-Value APIs Native Block APIs NVMe compliant Economy of Scale Performance: TCO/IOPS Capacity: TCO/TB Reliability: TCO/Downtime
13. Pliops XDP Explained 13
14. System Integration Overview NVMe/Block Driver or NVMe Driver CPU / OS (or DPU) NW Block Driver Pliops Agent 14 QLC TLC Optane™
15. High Performance Drive Fail Protection w/ Fast Rebuild Compression Line-Rate Merge Pack Sort Index Garbage Collection Key-Value Storage Engine KV Library API RocksDB, NVMe-KV Compatible NVMe Block Interface With Advanced Thin Provisioning 15 Hardware virtualization for QoS per Volume* * Next Generation Capabilities SDXI, TP-4091* Storage & Analytics Acceleration ARM Core Complex* RTL XDP Functional Flow Sequentialize 100% Sequential Writes to SSDs maximizing performance, endurance
16. Application IO Amplification Challenge 4K 40x IO amplification 100 Bytes IO TLC indirection unit 1 2 • Impacts Network, Storage, SSD, CPU – Must Overprovision for This Extra Data Transfer and Processing • Improving IO Amplification consumes CPU, cache or storage space. 64K 64x IO amplification 1KB IO QLC indirection unit
17. Optimization Points and Tradeoff Relationships
18. 18 Tradeoffs in Action Pliops SW-Based Storage Engines
19. Storage Engines are Everywhere! InnoDB WiredTiger RocksDB MyISAM XtraDB Applications Packing/meta data mgt (AFA or advanced SDS) Drive Fail Protection Sequentialization (advanced AFAs only) DAS AFA SDS/DFS System resource consumption 30-70% CPU Consumption 2,000-10,000% IO Amplification 20%-500% Storage Capacity Amplification 10-100% CPU Consumption 300-600% Storage Capacity Amplification Analytics NoSQL RDBMS Storage Block Object DFS Storage Engines Storage Management Compression Block Physical More Than Half of System Resources Typically Devoted to Storage Engine and Services! NAND Management 15-50% Cost of SSD DRAM, Processor, NAND overprovisioning
20. Pliops XDP Accelerates Core Storage Functions InnoDB WiredTiger RocksDB MyISAM XtraDB Applications Storage Engines Storage Management DAS AFA SDS/DFS Analytics NoSQL RDBMS Storage Block Object DFS •Collapses layers into one optimal device​ •Removes 50%-500% of system resources •Brings IO Amplification to theoretical minimums - <1x to 3x max •Near universal applicability Block Physical Compression Packing/meta data mgt (AFA or advanced SDS) Drive Fail Protection Sequentialization (advanced AFAs only) NAND Management Enables extreme scaling for nearly all flash-based workloads
21. XDP Delivers Ultra-Consistent Performance Even with Full Drives 0 1 2 3 4 5 6 7 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Random Write Performance Used SSD Capacity Datacenter SSDs Pliops XDP
22. 0 0.2 0.4 0.6 0.8 1 1.2 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 512 1024 4096 16384 MBPS Random Read Performance at various object sizes SW read performance XDP read performance 0 0.2 0.4 0.6 0.8 1 1.2 0 200 400 600 800 1000 1200 1400 1600 1800 2000 512 1024 4096 16384 MBPS Random Write Performance at various object sizes SW write performance XDP write performance RocksDB Storage Engine Performance 22 Orders of magnitude performance acceleration with Pliops XDP
23. 0 500 1000 1500 2000 2500 3000 0 200000 400000 600000 800000 1000000 1200000 Microseconds QPS Latencies Determine Max Performance RocksDB (90% RR/10% RW) Average Latency Data centers can dramatically increase scale within existing constraints RocksDB Pliops XDP
24. Storage Capacity Benefits with Pliops 24 RAID 10 Setup User Capacity: 10.7TB 3.84 TB x 8 SSDs RAID 10 TBs Pliops Setup User Capacity: 67TB 3. 84TB x 8 SSDs TBs 11TB usable 67TB Usable 6x increase
25. Industry Working Around Problems - Expensively https://www.mongodb.com/docs/manual/administration/production-notes/ https://docs.datastax.com/en/landing_page/doc/landing_page/systemChecks.html RAID 5/6 has been designed out due to poor performance Cassandra Space Amp requires 2-3x more storage than user data for compaction phase “For optimal performance in terms of the storage layer, use disks backed by RAID-10. RAID-5 and RAID-6 do not typically provide sufficient performance to support a MongoDB deployment.” RAID Check if RAID is used. Don’t use RAID-5, RAID-6, or their variants, such as RAID- 50/RAID-60, which has worse performance characteristics. Compaction strategy SizeTieredCompactionStra tegy(STCS) Minimum requirements Sum of all the SSTables compacting must be smaller than the remaining disk space. Note: Worse case: 50% of free disk space. This scenario can occur in a manual compaction where all SSTables are merged into one giant SSTable. https://docs.datastax.com/en/oss-planning/doc/planning/ossCapacityPlanning.html
26. 26 Pliops XDP vs HW RAID 5/6 Benefits TLC SSD: 8 Samsung PM9A3 With Pliops Performance Impact from Read/Modify/Writes on write 5x+ Performance Benefits for TLC and QLC SSDs Longer Rebuild Cycles– Rebuild Entire Drive Faster Rebuilds – Rebuild Only Used Data Impacts SSD Useful Life Extends SSDs Life Cycle Including QLCs No built in Data Compression Up to 6x Increased Storage Capacity* Hardware RAID 5/6 Customer Needs Rebuild Duration Performance *Includes RAID efficiency gains SSD Endurance SSD Capacity Expansion
27. Customer Use Cases 27
28. Top 5 SaaS Provider XDP Benefits S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 ... S23 S24 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S23 S24 ... 20 User Instances 66TB Usable, XDP DFP Zero Server Failures/Year/ across all servers Full Rack Software RAID vs Pliops XDP Customer Benefits Current Setup Pliops Setup 15.84TB x 2 SSDs Db1 Dbn ... 15.84TB x 4 SSDs Db1 Dbn ... 15 User Instances 21 TB Usable, RAID 0 600 Server Failures/Year/ 10K servers 3.2x ↑ Higher Capacity Zero Eliminated >600 SSD Related Failover Events 58% ↓ Reduction in TCO/TB User Data Improved Customer Experience/ Satisfaction 33% ↑ Increase Multitenancy Database Instances based on 3 years
29. Top 5 CSP Flash Storage Service XDP Benefits Accelerated High-Capacity EBS Customer Benefits S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 Software-Defined Storage File System 7.6TB x 24 SSDs Current Setup Pliops Setup 15.36TB x 22 SSDs Hardware Accelerated Software-Defined Storage 10+2 Servers 2 Erasure Coding 1.5PB Usable > 1500 Rebuilds /Year 10+2 Servers 2 Erasure Coding 4.3PB Usable XDP Drive Fail Protection 0 SSD-Related Failures/Year 3x ↑ Increase in Effective Storage Capacity 1.5x ↑ Improved Endurance Using QLC vs. SW TLC 247% ↓ Reduced TCO/TB Reduced Carbon Footprint Considerably Isolate Drive Failures from Cluster Performance Impact (FTT n+’1’) based on 3 years
30. RAID: XDP Storage & Performance Benefits Pliops Setup 7.68TB x 8 SSDs Cost Per TB: $400 User Capacity: 111TB 30 RAID 5 Setup 7.68 TB x 8 SSDs User Capacity: 38TB Cost Per TB: $1146 XDP Capacity Benefits 3x ↑ Higher User Capacity 65% ↓ Reduction in TCO/TB User Data Performance Benefits 5x ↑ Average Higher Performance 74% ↓ Reduction in TCO/IOPS based on 3 years
31. 31 In-Memory vs. Large Data Sets on Flash Redis on DRAM 100s of GBs Redis on Flash RocksDB 100s of GBs TBs Redis on XDP Pliops RocksDB API 100s of GBs TBs Performance X X ✓ Latency X X ✓ Data Set Size X X ✓ RA/WA/SA - X ✓
32. Redis In-Memory - TCO Advantage 22 Instances of Redis on Flash + Pliops RocksDB API S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 Current Setup Redis on Flash + XDP Pliops Setup Redis on Flash vs Redis with Pliops XDP Customer Benefits Redis In-Memory 3.84TB x 4 SSDs 22 Instances of Redis In-Memory+ RocksDB 32 based on 3 years 13x ↑ Higher Capacity Extend business insights from few weeks to several months 92% ↓ TCO/Reduction Reduction
33. Redis on Flash TCO Advantage 22 Instances of Redis on Flash Pliops RocksDB API 959 KIOPS, 19.9ms 99.99% Latency S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 22 Instances of Redis on Flash 145 KIOPS, 79ms 99.99% Latency Current Setup Pliops Setup Redis on Flash vs Redis with Pliops XDP Customer Benefits 3.84TB x 4 SSDs 3.84TB x 4 SSDs 33 based on 3 years 7x ↑ Higher Performance 86% ↓ TCO/IOPS Reduction 4x ↓ Lower 4 9’s Latency Improved Customer Experience/ Satisfaction 5.8x ↑ Improved Endurance
34. 50% ↓ Reduced TCO CPU/Mem Usage X ✓ ✓ Storage/Compute Scaling X ✓ ✓ WA/RA/SA X X ✓ Network Utilization X X ✓ TCO X X ✓ KV Distributed Systems 34 based on 3 years Pliops Benefits 5x ↑ More throughput Storage Server RocksDB Network Network Pliops RocksDB API Compute Server KV Application RocksDB Compute Server KV Application Compute Server KV Application Storage Server
35. SmartBOF powered with XDP+DPU • Built-in drive fail protection with virtual hot capacity • Ultra-fast rebuild with minimal performance impact • Present multi-volumes with configurable characteristics • Built-in KV Storage Engine • Hot-add/remove capacity/drive • Improve data endurance (1.5x using QLC vs. SW TLC) • Increase effective storage capacity (~3x ↑) • Reduce TCO (40% ↓) Pliops Customer Benefits 35 JBoF/AFA Server JBoF/AFA EBoF/AFA Storage • DPU is running NMVe-oF target/offload and Pliostore • XDP owns all NVMe-oF devices with multi-volumes* • DPU as-host CPU; no need for expensive x86 BOM • Optimized with P-to-P; dual-port NVMe XDP DPU * Optimized data flow between XDP and DPU DPU GPU GDS
36. Customer and Partner Endorsements 36
37. Industry Support for Pliops “Pliops’ technology enables the ability to turn ‘dumb’ flash storage into ‘smart’ flash storage. According to our benchmarks, the Pliops Storage Processor is unique in that it is able to increase performance, improve compression and reduce write amplification.” David Dixon VP, Data Center NAND Division Peter Zaitsev CEO Rob Davis VP of Storage Technology “We agree with the Pliops vision that data intensive applications will be critical to tomorrow’s innovations, so it’s imperative to architect the computing world for what’s coming. The partnership with Pliops is a key part of our strategy to continue to lead the way through the AI revolution. Pliops gives customers and important tool they can use to accelerate, and cost reduce SSD based applications” “With it’s XDP product, Pliops is leading the revolution in a way data is processed and stored in the datacenter. Inspur Systems is partnering with Pliops in delivering robust, performance-optimized, solutions. From Cloud Services Providers to marquee enterprise customers, together with Pliops we deliver an optimized stack to address data acceleration, while also dramatically enhancing media reliability and endurance as we witness a tidal shift towards QLC NAND-based storage” Dolly Wu VP of Cloud “Simply put, Pliops XDP unleashes the full potential of our QLC 3D NAND SSDs. This opens the door for all use case, delivering performance and endurance greater than what can be achieved using more expensive alternatives such as TLC SSDs. By extending the value of QLC for all existing applications and multiplying reliability, performance, capacity and efficiency, Pliops XDP enables more scaling for SSD investments” 37
38. Public Customer Support for Pliops Takeshi Hasegawa Engineering Manager “Thanks to Pliops’ technology, we are able to realize higher performance from our investments and support more concurrent users. Using Pliops Storage Processor provides a superior architecture, with high performance and space savings all with drive failure protection. This is the best product for server-side flash that I have seen in 10 years.” Pliops XDP will enable us to extend the storage capacity used by Redis, our in-memory data service. This means our engineers can analyze months of data instead of days, improving their insights on product development and validation. Anshuman Singh Director of SW Development 38
39. Building More Sustainable Data Centers 39
40. 40 WW Data Center Carbon Emissions https://www.eesi.org/papers/view/fact-sheet-the-growth-in-greenhouse-gas-emissions-from-commercial-aviation https://enabledenergy.net/blog/did-you-know-these-data-center-facts/ 7,500 data centers WW 5.75M servers added/year 40% of cost for P&C WW Data Centers Carbon Emissions Global Airlines Industry =
41. • Without Pliops, global datacenter energy consumption grows to 14% of global CO2 emissions • Pliops enables power efficient data centers to accomplish green data center goals • Enables 39% more datacenter compute / storage with 45% less CO2 per server/storage unit • Additionally, reduces emissions and waste during manufacturing and decommissioning servers • Expands global access to advanced internet services Building Next Generation Data Centers with Pliops 41 Equal to removing 46M cars through 2040 Accelerates Carbon Neutral Timeline Massive Data Center Power Savings Reduce 214 Megatons CO2/yr by 2040 PLIOPS XDP
42. Thank You! @Web @Twitter @LinkedIn 42