Some early results for SiCortex machines

Some early results for SiCortex machines

John F. Mucci

Founder and CEO, SiCortex, Inc.

Lattice 2008

The Company

Computer Systems company building complete, high processor count, richly interconnected, low power Linux computers

Strong belief (and now some proof) that a more efficient HPC computer can be built from the silicon up.

Around 80 or so really bright people, plus me.

Venture funded, based in Maynard, Massachusetts, USA

http://www.sicortex.com, for whitepapers and tech. info

What We Are Building

A family of fully-integrated HPC Linux systems delivering best-of-breed:

Total Cost of Ownership

Delivered Performance:Per Watt

Per Square Foot

Per Dollar

Per Byte/IO

Usability and deployability

Reliability

SiCortex Product Family

105 Gflops48 GB

6.5 GB/s I/O200 Watts

0.95 Teraflops864 GB

30 GB/s I/O2.5 KWatts

8.55 Teraflops7.8 Terabytes216 GB/S I/O20.5 Kwatts

2.3m2

2.14 Teraflops1.8 Terabytes68 GB/S I/O5+ KWatts

Make it Green, Don't Paint it Green

Through increasing component density and integration ~= Performance

~= Reliability

~= 1/Power

Innovate where it counts!Single core performance and architecture

ScalabilityMemory, Network, I/O

ReliabilityNetwork, on-chip and off-chip ECC, software to recover

Software usability

Buy, borrow, the rest...

6

The SiCortex Node Chip

QuickTime™ and a decompressor

are needed to see this picture.

7

27-Node Module

3x 8-lane PCIe Modules

27x Node

54x DDR2 DIMM

2x Gigabit Ethernet

Fibre Channel10 Gb EthernetInfiniBand

Compute: 236 GF/secMemory b/w: 345 GB/secFabric b/w: 78 GB/secI/O b/w: 7.5 GB/sec

Fabric Interconnect

8

35

34

33

32

31

30

29

28

27

26 25 24 23 22 21 20 19 18

0 1 2 3 4 5 6 7 8

9

10

11

12

13

14

15

16

17

Network and DMA Engine

NetworkUnidirectional, 3 unique routes between any pair.3 in, 3 out, plus loopback, each 2GB/sFully passive, reliable, in order, no switch

DMA Engine100% user level, no pinningSend/Get/Put semanticsRemote Initiation

MPI: 1.0us, 1.5GB/sec

9

Standard Linux/MPI Environment

Integrated, Tuned and TestedOpen Source:• Linux• GNU C, C++, Fortran• Cluster file system (Lustre)• MPI libraries (MPICH2)• Math libraries• Performance tools• Management software

SiCortex:• Optimized compiler• Console, boot, diagnostics• Low-level communications

libraries, device drivers• Management software• 5 Minute boot time

Licensed:• Debugger, Trace Visualizer

GNU

MPI

Libraries

gentooLinux

QCD: MILC and su3_rmd

A widely used Lattice Gauge Theory QCD simulation for:

Molecular dynamics evolution, hadron spectroscopy, matrix elements and charge studies

The ks_imp_dyn/su3_rmd case is a widely studied benchmark.

Time tends to be dominated by the Conjugate Gradient

http://physics.indiana.edu/~sg/milc.html

http://faculty.cs.tamu.edu/wuxf/research/tamu-MILC.pdf

MILC su3_rmd Scaling (Input-10/12/14)

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

1 10 100 1000 10000

# Cores

Slowdown

v10v12v14AMD64/IBBGL

12

Understanding what might be possible

SiCortex system is new; compared to 10+ years of hacking and optimizationSo we took a look at a test suite and benchmark provided by Andrew Pochinsky @ MITLooks at the problem in three phases

What performance do you get running from L1 cacheWhat performance do you get running from L2And from main memoryVery useful to see where cycles and time are spent.

And gives hints about what compilers might do and how to restructure codes.

13

So what did we see?

By select hand coding of Andrews code we have seen:

Out of L1 cache 1097 Mflops Out of L2 cache 703 MflopsOut of Memory 367 Mflops

Compiler is improving each time we dive deeper into the code.But we’re not experts on QCD, could use some help.

What conclusion might we draw

Good communications makes for excellent scaling (MILC)

Working on single node performance tuning (on Pochinsky code) gives direction on performance and insight for compiler.

DWF formulations have higher computation/communications ratio. And we do quite well. Will do even better with formulations that have increased communications.

15

SiCortex and Really Green HPC

Come downstairs (at the foot of the stairs) and take a look and give it a try.

It’s a 72 processor (100+ Gflop) desktop system using ~200 watts. Totally compatible with its bigger family members. Up to 5832 processor system.

More delivered performance per square foot, per dollar, and per watt

16

Performance Criteria for the Tools Suite

•Work on unmodified codes•Quick and easy characterization of:

– Hardware utilization (on and off-core)– Memory– I/O– Communication– Thread/Task load balance

•Detailed analysis using sampling•Simple instrumentation•Advanced instrumentation and tracing•Trace-based visualization•Expert access to PMU and perfmon2

17 Proprietary and Confidential

• papiex - Overall application performance • mpipex - MPI profiling• ioex – I/O profiling• hpcex - source code profiling• pfmon - highly focused instrumentation• gptlex – dynamic call path generation• tauex - automatic profiling and visualization• vampir - parallel execution tracing

• gprof is there too (but is not MPI-aware)

Application Performance Tool Suite

18

For fun (and debugging)

19

For fun (and debugging)

20

Thanks