Make It So ! A Tale of Two Needs Dr. Robert W. Wisniewski Chief Software Architect Extreme Scale Computing Senior Principal Engineer, Intel Corporation.

Make It So !

A Tale of Two Needs

Dr. Robert W. WisniewskiChief Software Architect Extreme Scale ComputingSenior Principal Engineer, Intel Corporation

Smoky Mountains ConferenceSeptember 4, 2014

Copyright © 2014 Intel Corporation. All rights reserved.

1

Legal Disclaimer

Results have been estimated based on internal Intel analysis and are provided for informational purposes only. Any difference in system hardware or software design or configuration may affect actual performance.

Intel, processors, chipsets, and desktop boards may contain design defects or errors known as errata, which may cause the product to deviate from published specifications. Current characterized errata are available on request.

Intel, Intel Xeon, Intel Core microarchitecture, and the Intel logo are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries.

*Other names and brands may be claimed as the property of others

Copyright © 2014, Intel Corporation. All rights reserved.

Motivation for this Talk

3

Your Supercomputer 25 Years Ago

hardware

kernel

compiler

Application

C O M P U T E | S T O R E | A N A L Y Z E

Cray XC30 Software Stack

4

5

Motivation for this Talk

• Building Web Pages• Recent PhD talk

• Turtles all the way down

• HW and SW getting more complex

6

Motivation for this Talk

• Building Web Pages• Recent PhD talk

• Rabbits all the way down

• HW and SW getting more complex

7

1701D

• Supposition: analogous to dark silicon, let the “free” cycles do the work

8

Make It So

Before Proceeding

• Not for everyone– Consider cut through high performance

• Good for new applications– Biological for example– Enables new ways of approaching the problem

• Delivering on the needs of the next machine, the next next machine, and the exascale machine

9

10

Exa

Peta

ZetaPEZExascale is only a point on the continuum

11

When investigations began• Challenges too great with current SW• Need all new OS, compiler, language…

Others advocated• Enhance capability of

existing

• Hard, drive evolutionary approach

OR

??

Extreme-Scale Software Challenge

Revolutionary versus Evolutionary

12

• Which one ?

Revolutionary Only

13

Imagine vendors telling their customers throw out everything you’ve done over the last 20+ years.Leverage tremendous investment in Intel Architecture ecosystem.

Evolutionary Only

14

But there are serious challenges getting to exascale. Drive new innovations and invigorate the x86 ecosystem.

The Real Extreme-Scale Software Challenge

15

• The real challenge in moving software to extreme scale, and therefore the real solution, will be figuring out how to incorporate and support existing computation paradigms in an evolutionary model while simultaneously supporting new revolutionary paradigms.

AND

1701D

• Supposition: analogous to dark silicon, let the “free” cycles do the work

16

Make It So

Maybe not so far fetched, done this a bit already

• Compilers

• Demand paging– No– Yes

17

mOS: multiple Operating System

18

• Run multiple OSes on node simultaneously• Linux API with LWK performance

• Kernel is configured for the application– Provides compatibility and performance/scalability/reliability

ComputeNode

Linux

I/O daemon

OSNode

N:1

LWKOSLinux

Application

HPC System

Platform Softwarepartitioned

glibc/shim

AppAssist

tools

Global OSResilienceDynamic

• Resolve as many faults as possible– Extra nodes swapped in automatically (proactively if possible)

– NVRAM accessible independent of node– Automatic checkpoint of DRAM to NVRAM

– For unresolvable provide fault information to application

19

CN CN

CN CN

CN CN

Linux API

Argo Hobbes

OSN

OSN

SMC

HPC

netw

ork

nvram

CN

SMC

SMC

Runtimes and Libraries

20

Effort Pyramid

Application (100+)

Runtime/Lib (10+)

HWLL SW(1+)

task i

datain

controlin

data control

task ntask n

task n

core core core core

task n

runtime• Over commit– Adaptive, asynchronous,

load balance, fault tolerance• Building block approach

Data Management for Big Data

• Smooth and automatic representation between– Application data structure in memory– Representation and access to NVRAM– Storage to disk

• Moving compute to data• Application makes system call

– make_permanent(*data), make_durable(*data)

21

main()

A[100][100][100];

graph_node { int value; edge e1;} RAM nvram

main()

22

It was the Best, It was the Worst of Times

• It was evolutionary, it was revolutionary

• It was research, it was development

Weaving the Two Tales (Needs) Together

– This is possible

23

Research

/Out-o

f-Box

Development/Evolutio

n & Revolutio

nSingle Unified Project

Time

Valu

e

Fundamental Discovery to Gain

Fundamental Insights

Technical Computing Continues Its Rapid Growth

Source: IDC: Worldwide Technical Computing Server 2013–2017 Forecast; Other brands, names, and images are the property of their respective owners.

Governments & Research Commercial/Industrial New Users – New Uses

Business Transformation

Big Data Analytics Enabling Data Driven Science

BetterProducts

Faster Timeto Market

ReducedR&D

From Diagnosis

to personal-

ized treatment

quicklyGenomics Clinical

Information

HPC: Transforming the world of data and information into KNOWLEDGE

“My goal is simple. It is complete understanding of the universe, why it is as it is and why it exists at all”

Stephen Hawking

Conclusion

– The Best of Both Worlds– Leverage PEZ(Y) cycles to provide higher level and more

productive abstractions to applications– We need to build tomorrow’s turtles/rabbits

– We will get to extreme scale (PEZ) by figuring out how to incorporate existing computation paradigms in an evolutionary model while simultaneously supporting new revolutionary paradigms

25

1701D

AND