Transcript Benchmarks_Parallel - the Academic & Research Wing of SERC

Benchmarks for Parallel
Systems
Sources/Credits:
“Performance of Various Computers Using Standard Linear Equations
Software”, Jack Dongarra, University of Tennessee, Knoxville TN,
37996, Computer Science Technical Report Number CS - 89 – 85, April
8, 2004, url:http://www.netlib.org/benchmark/performance.ps
http://www.top500.org
FAQ: http://www.netlib.org/utk/people/JackDongarra/faq-linpack.html
Courtesy: Jack Dongarra (Top500)
http://www.top500.org
The LINPACK Benchmark: Past, Present, and Future,
Jack Dongarra, Piotr Luszczek, and Antoine Petitet
NAS Parallel Benchmarks. http://www.nas.nasa.gov/Software/NPB/
LINPACK (Dongarra: 1979)
 Dense system of linear equations
 Initially used as a user’s guide for
LINPACK package
 LINPACK – 1979
 N=100 benchmark, N=1000
benchmark, Highly Parallel Computing
benchmark
LINPACK benchmark
 Implemented on top of BLAS1
 2 main operations – DGEFA(Gaussian
elimination - O(n3)) and DGESL(Ax =
b – O(n2))
 Major operation (97%) – DAXPY: y =
y + α.x
 Called n3/3 + n2 times. Hence 2n3/3
+ 2n2 flops (approx.)
 64-bit floating point arithmetic
LINPACK
 N=100, 100x100 system of equations. No change in
code. User asked to give a timing routine called
SECOND, no compiler optimizations
 N=1000, 1000x1000 – user can implement any code,
should provide the required accuracy: Towards Peak
Performance (TPP). Driver program always uses 2n3/3
+2n2
 “Highly Parallel Computing” benchmark – any
software, matrix size can be chosen. Used in Top500
 Based on 64-bit floating point arithmetic
LINPACK
 100x100 – inner loop optimization
 1000x1000 – three-loop/whole program
optimization
 Scalable parallel program – Largest
problem that can fit in memory
 Template of Linpack code




Generate
Solve
Check
Time
HPL (Implementation of HPLinpack
Benchmark)
HPL Algorithm
• 2-D block-cyclic data distribution
•Right-looking LU
•Panel factorization: various options
- Crout, left or right-looking recursive variants based on
matrix multiply
- Number of sub-panels
- recursive stopping criteria
- pivot search and broadcast by binary-exchange
HPL algorithm
 Panel broadcast:
-
 Update of trailing
matrix:
- look-ahead
pipeline
 Validity check
- should be O(1)
Top500 (www.top500.org)
 Top500 – 1993
 Twice a year – June and November
 Top500 gives Nmax, Rmax, N1/2,
Rpeak
India and Top 500
Rank
Site
111
Geoscience (B)
India
204
System
Vendor
Processors
Rmax
Rpeak
BladeCenter HS20
Cluster, Xeon
EM64T 3.4 GHz Gig-Ethernet
IBM
1024
3755
6963.2
Semiconductor
Company (L)
India
eServer, Opteron 2.6 GHz,
GigEthernet
IBM
1024
2791
5324.8
231
Semiconductor
Company (K)
India
xSeries x336 Cluster Xeon
EM64T 3.6 GHz Gig-Ethernet
IBM
730
2676.88
5256
293
Institute of Genomics
and Integrative
Biology
India
Cluster Platform 3000
DL140G2 Xeon 3.6
GHz Infiniband
Hewlett-Packard
576
2156
4147.2
NAS Parallel Benchmarks - NPB
Also for evaluation of Supercomputers
A set of 8 programs from CFD
5 kernels, 3 pseudo applications
NPB 1 – Original benchmarks
NPB 2 – NAS’s MPI implementation. NPB
2.4 Class D has more work and more I/O
 NPB 3 – based on OpenMP, HPF, Java
 GridNPB3 – for computational grids
 NPB 3 multi-zone – for hybrid parallelism





NPB 1.0 (March 1994)
 Defines class A and class B versions
 “Paper and pencil” algorithmic
specifications
 Generic benchmarks as compared to
MPI-based LinPack
 General rules for implementations –
Fortran90 or C, 64-bit arithmetic etc.
 Sample implementations provided
Kernel Benchmarks
EP – embarrassingly parallel
MG – multigrid. Regular communication
CG – conjugate gradient. Irregular long distance
communication
 FT – a 3-D PDE using FFT. Rigorous test of long distance
communication
 IS – large integer sort
 Detailed rules regarding
- brief statement of the problem
- algorithm to be practiced
- validation of results
- where to insert timing calls
- method for generating random numbers
- submission of results



Pseudo applications / Synthetic
CFDs
 Benchmark 1 – perform few iterations
of the approximate factorization
algorithm (SP)
 Benchmark 2 - perform few iterations
of diagonal form of the approximate
factorization algorithm (BT)
 Benchmark 3 - perform few iterations
of SSOR (LU)
Class A and Class B
Class A
Sample Code
Class B
NPB 2.0 (1995)
 MPI and Fortran 77
implementations
 2 parallel kernels (MG,
FT) and 3 simulated
applications (LU, SP,
BT)
 Class C – bigger size
 Benchmark rules –
0%, 5%, >5% change
in source code
NPB 2.2 (1996), 2.4 (2002), 2.4
I/O (Jan 2003)
 EP and IS added
 FT rewritten
 NPB 2.4 – class D and rationale for class D
sizes
 2.4 I/O – a new benchmark problem based
on BT (BTIO) to test the output capabilities
 A MPI implementation of the same (MPI-IO)
– different options using collective buffering
or not etc.