Hartmut Schultze May 15th, 2019 - Hewlett Packard …...graph inference 100x faster Memory-Driven Computing Software Defined - dynamically configure right-sized solutions in real time

– Hartmut Schultze

– May 15th, 2019

2

The DZNE

One center, several locations

75

433

755

9331031

1120

2009 2011 2013 2015 2017 2018

Number of staff

81 research groups

Scientists from

54 countries

34 % Internationalscientists

4019 Publications(till Dec. 2017)

Development cycle in systems medicine

+ Knowledge+ Understanding

+ Therapy+ Risk minimization

Patients

High Throughput

Big DataAnalytics

Models,diagnostic

tools

Validation

Clinical Data Laboratory

Imaging

Computing is central

I/O

Copper

DRAM

DRAM

Copper

SoC

SoC

SoC

SoC

Memory

Mem

ory

Memory

Mem

ory Memory-Driven

ComputingProcessor-Centric

Computing

Persistent Memory

+Fabric

Classical vs. Memory-Driven Computing architecture

Today’s architecture

is constrained by the CPU

DDR

Ethernet

PCI

If you exceed what can be connected to one CPU, you need another CPU

Memory-Driven Computing:

Mix and match at the speed of memory

SATA

Photonic fabrics destroy distance

Photonics

Hundreds of racks can behave as a single server

= 4096 yottabytes (=4096*1024 Bytes)292

bytes

use case - genomic data analysis

Can we make the processing of the raw genome data faster with Memory-Driven Computing?

▪ Using Superdome X

▪ Memory-Driven Computing -Architecture

▪ Small Change in Code

Near optimal RNA-Seq quantification

13secprocessing time reduced from22 minutes

112xincrease in analytics speed removes research bottlenecks

60%energy reductioncuts research costs

Change Mindset

–Think of having ubiquitous memory

–Think of access to memory having no interference from other nodes

–Rethink code

–Change mindset on approach

Transform performance with Memory-Driven programming

14

In-memory analytics

15xfaster

New algorithms Completely rethinkModify existing

frameworks

Genomics

100xfaster

Financial models

10,000xfaster

Large-scale

graph inference

100xfaster

Memory-Driven ComputingSoftware Defined - dynamically configure right-sized solutions in real time

Processors

Microprocessor/

System on Chip (SoC)

Memorydata storage, volatile and non-

volatile

Storage Class Memory/

Persistent Memory

Non-volatile, fast

Flash

Non-volatile, slow

DRAM

Volatile, fast, high energy

consumption

AcceleratorsSpecialized processors designed to perform

some functions more efficiently

GPU

FPGA

Array of logic gates that can be hardware-

programmed to fulfill user-specific tasks

Dot Product Engine/

Neuromorphic Computing

high performance per-watt task-specific

analog accelerator computing using

Memristor Arrays

EDGE DEVICE

HYPERCONVERGED

SYSTEMExascale

Eliminate data movement via shared memory and reduce energy consumption

Fabric

CPUs Accelerators

Memory technologies I/O

– High bandwidth

– Low latency

– Advanced workloads & technologies

– Scalable from IoT to exascale

– Compatible

– Economical

– Supports electrical or optical interconnects

– Open standard

– Security built-in at the hardware level

Gen-Z: new open interconnect protocolKey enabler of the Memory-Driven Computing open architecture

FPGA

GPUSoC ASICNEURO

Memory

Memory

Network Storage

Direct Attach, Switched, or Fabric Topology

NVM NVM NVM

SoC

Memory

Open consortium with broad industry support

Consortium Members

System OEM

Cisco

Cray

Dell EMC

HPE

Huawei

Lenovo

NetApp

Nokia

Yadro

CPU/Accel

AMD

Arm

Cavium

IBM

Qualcomm

Xilinx

Mem/Storage

Everspin

Micron

Samsung

Seagate

SK Hynix

Smart Modular

Spintransfer

Toshiba

WD

Silicon

Broadcom

IDT

Mellanox

Microsemi

IP

Avery

Cadence

Intelliprop

Mobiveil

PLDA

Eco/Connect

Alpha Data

AMP

FIT

Hirose

Jess Link

Keysight

Lotes

Luxshare

Molex

Senko

TE

Tyco

Software

Redhat

Vmware

Govt/Univ

ETRI

Oak Ridge

Simula

UNH

Yonsei

Technology

Service Provider

Google

Microsoft

HPE Superdome FlexThe undisputed leader in in-memory computing

18

Designed with Memory-Driven Computing principles

Turn data into actionable insights in real time

– Unparalleled scale 4-32 sockets, 768GB-48TB+ memory

– Highly expandable for growth; ultra-fast fabric

Keep pace with evolving business demands

– Unique modular 4-socket building block

– Open management and hard partitioning for hybrid IT consumption

Safeguard mission-critical workloads

– Proven Superdome RAS with 99.999% single system availability

– Mission critical expertise with HPE Pointnext services

Pushing the limits

– Software-Defined Scalable Memory project in development to extend to 96TB of

composable memory across multiple racks

Software-Defined Scalable Memory on HPE Superdome FlexDeploying in our Memory-Driven Computing Sandbox

–

–

–

–

[email protected]