E8 Storage, Proprietary and Confidential About E8 Storage

E8 – Serving Data Hungry Application

October 2018 Ziv Serlin - VP Architecture & Co-founder Presenter

In E8 Mr. Serlin was involve on architect E8 product for the 1st two years, recently Mr. Serlin is focusing on the extensive engagement with E8 end customers exploring, helping and defining with the customers their E8 deployment models, In Ziv Serlin addition to that Mr. Serlin lead the strategic engagement of E8 HW partners. An Co-Founder & VP Architecture expert in designing complex storage systems, Mr. Serlin has extensive experience as IBM XIV, Primary Data, Intel a system architect at Intel and was the HW R&D manager of IBM/XIV product (scale out high-end enterprise block storage system). He earned BSC in Computer Engineering at the Technion. He hold storage patent from his work at IBM.

• Founded in November 2014 by storage industry veterans from IBM-XIV • Leading NVMe over Fabrics solution in the market • In production with customers in U.S. and Europe, expanding now to Asia • Awarded 10 patents (granted) + 4 pending for E8 architecture • World-wide Team: • R&D in Tel-Aviv • Sales & marketing in U.S., Europe and Asia • Flash Memory Summit 2016 & 2017 Most Innovative Product Award

5% Business volume grows 5% a year

60% The size of their database grows 60% a year

100% Performance load on that database grows 100% year

Big Data Analytics Image Recognition Real-time Security AI/ML Video 4K Post-Prod

E8 Storage Accelerates Data Hungry Applications

Local SSDs today achieve latency 10x faster than all-flash arrays • “The DevOps Problem” Local SSD AFA • Things that work on laptops become 10x slower on the production infrastructure 0.1ms 1ms

• “The islands of storage problem” ??? • Local SSDs in servers mean inefficient capacity utilization, no sharing of SSD data • Local SSDs couple storage and compute • Server purchasing requires upfront investment in SSDs

• Not enough performance • Traditional all-flash arrays (SAN/NAS) get 10%-20% of the potential performance of NVMe SSDs • Classic “scale-up” bottleneck • Dual controller bottleneck • All I/O gated by controller CPU • Switching the SSDs from SAS to NVMe cannot alleviate the controller bottleneck First gen architectures cannot unlock the full performance of NVMe

• By attaching SSDs over the network, rather 4 Year Cost of SSDs in a Data-Center than using local SSDs in each server, Customer is able to: $80 • Reduce total acquisition cost of SSDs by 60% $60 • Defer capacity purchases until needed $40 • Pay as you grow – at lower future prices $20 $30 • Improve capacity utilization of SSDs to 90% $26 $20 • Shared volumes vs local replicas $12 • Easily add capacity only as needed

Moving to shared NVMe storage delivers strong TCO for large scale customer deployments

10M AFA with 24 SSDs 40 1000 Single NVMe SSD E8 24 NVMe SSDs

See the Demo! 100 120 750K 2.4 3.1 300K

IOPS (@4K read) Read/Write Bandwidth (GB/s) Read Latency (us)(@4K)

Communication protocol designed specifically for flash storage • High performance, low latency • Efficient protocol with lower stack overhead • Exponentially more queues / commands than SAS • Parallel processing for SSDs vs serial for HDDs • Support for fabrics (NVMe-oF™) • Originally designed for PCIe (internal to servers) • Expands support for other transport media • RDMA Based: RoCE, iWARP, Infiniband I/O Queues Commands per Queue • Non-RMDA: FC, TCP SAS 1 256 • Maintains NVMe protocol end to end NVMe 65,535 64,000

A new architecture built specifically for high performance NVMe

• Unleash the parallelism of NVMe SSDs • Direct drive access for near line rate performance • Separation of data and control paths; no controller bottleneck • E8 Agent offloads up to 90% of data path operations • Simple, centralized management • Intuitive management GUI for host / volume management • E8 Agents auto-discover assigned LUNs • Scalable in multiple dimensions • Up to 126 host agents per E8 Controller • Up to 8 Controllers per host  2PB in single management

No single point of failure anywhere in the architecture

• Host agents operate independently Host Servers with E8 Host Agents • Failure of one agent (or more) does not affect other agents • Access to shared storage is not impacted • RAID-6/RAID-5/RAID-10 data protection • Network multi-pathing • Enclosure high availability • Option 1: HA enclosure + dual-ported SSDs • Option 2: Cross-enclosure HA + single-ported SSDs

Shared NVMe as a fast tier for parallelizing genomic processing

• "We were keen to test E8 by trying to integrate it with our Univa Grid Engine cluster as a consumable resource of ultra-performance scratch space. Following some simple tuning and using a single EDR link we were able to achieve about 5GB/s from one node and 1.5M 4k IOPS from one node. Using the E8 API we were quickly able to write a simple Grid Engine prolog/epilog that allowed for a user-requestable scratch volume to be automatically created and destroyed by a job. The E8 box behaved flawlessly and the integration with InfiniBand was simpler than we could have possibly expected for such a new product." • - Dr. Robert Esnouf, Director of Research Computing Oxford Big Data Institute + Wellcome Center for Human Genetics

From 10 hours per genome to 1 hour for 10 genomes!

Shared NVMe Accelerates Deep Learning Cost ($/GBu) GPU Farm: Nvidia DGX-1 • A GPU cluster requires 0.5PB-1PB of shared fast storage • Up to 8 GPUs per node • GPFS Client + E8 Agent run on • But GPU servers have no real x86 within GPU Server estate for local SSDs… • Up to 126 GPU nodes in cluster • E8 Storage provides concurrent Pure Storage IBM GPFS + ESS E8 + GPFS access for 1000 (!) GPUs per FlashBlade cluster Mellanox 100G IB Images per second, per GPU node • 10x Performance of 3000 (ResNet-50 Image Recognition Training) Pure Storage FlashBlade 2500 Shared NVMe Storage 2000 • 4x Performance of IBM ESS SSD • E8-D24 2U24-HA 1500 Pure Storage Appliances, for half the cost IBM GPFS + ESS • Dual-port 2.5” NVMe Drives 1000 E8+GPFS • Up to 184TB (raw) per 2U 500 Patented Distributed RAID6 • 0 1 GPU node 10 GPU nodes 100 GPU nodes

Shared NVMe Accelerates AI/ML Workloads Cost ($/GBu) 7 6 • GPU Environment is Perfect for E8 5 • x86 processors on GPU nodes are free: run GPFS client + E8 agent 4 • GPU cluster typically connected with 100G IB or ROCE 3 2 • GPU nodes have no real estate for SSDs: requires external shared SSD solution 1 0 Pure Storage IBM GPFS + ESS Weka.IO (NVMe) E8 + GPFS • Highly Scalable, up to 1000 (!) GPUs in single cluster FlashBlade • 1-8 GPUs per node, up to 126 nodes per cluster • Scale number of NVMe drives and NVMe enclosures easily: 20TB – 2PB Epoch Time (Hours) 30

25 • 10x Performance of Pure Storage FlashBlade 20 • AI/ML data-sets incompatible with de-dup and compression 15 • Which also ruins performance… 10 5

0 • 4x Performance of IBM ESS SSD Appliances, for half the cost Pure Storage IBM GPFS + ESS Weka.IO (NVMe) E8 + GPFS • Open hardware architecture FlashBlade • No NSDs: Single hop from GPFS clients to SSDs through E8’s patented architecture • All network attach, no SAS cabling, no one-to-one mappings

Better Throughput

Better Throughput: E8 Storage processed more images per second than local SSD

Lower Latency

Client Client Client Client • Scalable to larger configurations E8 Agent • Can mix connectivity depending on requirements • Standalone pool NSD NSD E8 Agent E8 Agent • Shared LUNs • LROC • Non-shared LUNs (direct connect clients only) IB/ RoCE • HAWC

E8 MDS E8-D24 (Dual-port SSDs) E8 MDS

SSD SSD SSD SSD SSD SSD

• 0.57ms - Record breaking response time! • 45% lower ORT for the same builds • 8x lower latency on average • The power of Intel Optane • Ultra-low latency for data intensive

apps 8x lower latency! • More performance, less hardware

E8 Storage 24 NVMe SSDs 2U WekaIO 64 NVMe SSDs 8U Huawei 6800 F 60 SAS SSDs 14U

* As of SPEC SFS®2014_swbuild results published August 2018. SPEC SFS2014 is the industry standard benchmark for file storage performance. See all published results at https://www.spec.org/sfs2014/results/ 19 ©2018 E8 Storage, Proprietary and Confidential STAC®-M3™ Benchmark for High Speed Analytics

Best STAC-M3 Response Times (ms) • High performance with eXtremeDB® • Real-time analytics for tick database 1T.YRHIBID-2 • Designed for today’s financial systems 1T.WKHIBID*

• Accelerated with Intel Optane technology 10T.MKTSNAP* • Record breaking performance! 0 50 100 150 200 250 1T.WRITE.LAT2 • Faster response times in 5 of 17 operations* 17x Faster! 1T.NBBO • More consistent response times overall 10T.VOLCURV • Lowest overall execution time 100T.VWAB-12D-NO • More than 2.5x faster than competition! 0 5000 10000 15000 20000 E8 Storage Competitor A Competitor B

* Of the published, audited results on https://stacresearch.com/ as of May 2018. Graphs show the 2 closest competitors for overall results. 20 ©2018 E8 Storage, Proprietary and Confidential Roadmap: NVMe Takes Over the Data Center

Compute Networking Storage

NVMe tiers Smart NICs NVMe Over Fabrics • Performance tier (Intel • NICs with CPUs allow E8 • E8 Software is an enabler Optane, Samsung ZNAND) Storage to run without host for managed JBOFs • Capacity tier: TLC/QLC, Intel friction • E8 Agents can talk to ruler / Samsung m.3 • Support for VMware and standard commodity • E8 automates tiering and Windows NVMe-oF JBOFs resiliency scheme per tier

Data Center Technology Trends Open the Door for E8 Storage Architecture

Accelerate application performance with more processing power • Before • After • E8 Agent runs on host server user space • E8 Agent runs inside the Smart NIC • Typically requires 1-2 CPU cores • Uses CPU on Smart NIC, not host server • Supports Linux servers • Support for Windows and VMware

View Live Demo with Broadcom and Mellanox Smart NICs

PCIe SSD Centralized Storage Hyper- Aﬀordable Performance Reliability scalability 100% COTS

• For more information, please feel free to contact • Alfred Hui – [email protected] (Regional Director, Asia ) • +86-184-2013-7210 • Wechat connect: