Transcript PGAS-SIAMPP06

Partitioned Global Address Space
Languages
Kathy Yelick
Lawrence Berkeley National Laboratory
and UC Berkeley
Joint work with
The Titanium Group: S. Graham, P. Hilfinger, P. Colella, D. Bonachea,
K. Datta, E. Givelberg, A. Kamil, N. Mai, A. Solar, J. Su, T. Wen
The Berkeley UPC Group: C. Bell, D. Bonachea, W. Chen, J. Duell,
P. Hargrove, P. Husbands, C. Iancu, R. Nishtala, M. Welcome
PGAS Languages
1
Kathy Yelick
The 3 P’s of Parallel Computing
• Productivity
• Global address space supports construction of complex
shared data structures
• High level constructs (e.g., multidimensional arrays)
simplify programming
• Performance
• PGAS Languages are Faster than two-sided MPI
• Some surprising hints on performance tuning
• Portability
• These languages are nearly ubiquitous
Kathy Yelick, 2
Partitioned Global Address Space
Global address space
• Global address space: any thread/process may
directly read/write data allocated by another
• Partitioned: data is designated as local (near) or
global (possibly far); programmer controls layout
x: 1
y:
x: 5
y:
l:
l:
l:
g:
g:
g:
p0
p1
x: 7
y: 0
By default:
• Object heaps
are shared
• Program
stacks are
private
pn
• 3 Current languages: UPC, CAF, and Titanium
• Emphasis in this talk on UPC & Titanium (based on Java)
Kathy Yelick, 3
PGAS Language Overview
• Many common concepts, although specifics differ
• Consistent with base language
• Both private and shared data
• int x[10];
and
shared int y[10];
• Support for distributed data structures
• Distributed arrays; local and global pointers/references
• One-sided shared-memory communication
• Simple assignment statements: x[i] = y[i];
or
t = *p;
• Bulk operations: memcpy in UPC, array ops in Titanium and CAF
• Synchronization
• Global barriers, locks, memory fences
• Collective Communication, IO libraries, etc.
Kathy Yelick, 4
Example: Titanium Arrays
• Ti Arrays created using Domains; indexed using Points:
double [3d] gridA = new double [[0,0,0]:[10,10,10]];
• Eliminates some loop bound errors using foreach
foreach (p in gridA.domain())
gridA[p] = gridA[p]*c + gridB[p];
• Rich domain calculus allow for slicing, subarray, transpose and
other operations without data copies
• Array copy operations automatically work on intersection
data[neighborPos].copy(mydata);
intersection (copied area)
“restrict”-ed (non-ghost)
cells
ghost cells
mydata
data[neighorPos]
Kathy Yelick, 5
Productivity: Line Count Comparison
Fortran
Lines of code
2000
C
1500
1000
MPI+F
500
CAF
0
UPC
NPB-CG
NPB-EP
NPB-FT
NPB-IS
NPB-MG
Titanium
• Comparison of NAS Parallel Benchmarks
• UPC version has modest programming effort relative to C
• Titanium even more compact, especially for MG, which uses multi-d
arrays
• Caveat: Titanium FT has user-defined Complex type and cross-language
support used to call FFTW for serial 1D FFTs
UPC results from Tarek El-Gazhawi et al; CAF from Chamberlain et al;
Titanium joint with Kaushik Datta & Dan Bonachea
Kathy Yelick, 6
Case Study 1: Block-Structured AMR
• Adaptive Mesh Refinement
(AMR) is challenging
• Irregular data accesses and
control from boundaries
• Mixed global/local view is useful
Titanium AMR benchmarks available
AMR Titanium work by Tong Wen and Philip Colella
Kathy Yelick, 7
AMR in Titanium
C++/Fortran/MPI AMR
Titanium AMR
• Chombo package from LBNL
• Bulk-synchronous comm:
• Entirely in Titanium
• Finer-grained communication
• Pack boundary data between procs
• No explicit pack/unpack code
• Automated in runtime system
Code Size in Lines
C++/Fortran/MPI
Titanium
AMR data Structures
35000
2000
AMR operations
6500
1200
Elliptic PDE solver
4200*
1500
10X reduction
in lines of
code!
* Somewhat more functionality in PDE part of Chombo code
Elliptic PDE solver running time (secs)
PDE Solver Time (secs)
Serial
Parallel (28 procs)
C++/Fortran/MPI
Titanium
57
53
113
126
Comparable
running time
Work by Tong Wen and Philip Colella; Communication optimizations joint with Jimmy Su Kathy Yelick, 8
Immersed Boundary Simulation in Titanium
• Modeling elastic structures in an
incompressible fluid.
• Blood flow in the heart, blood clotting,
inner ear, embryo growth, and many more
• Complicated parallelization
• Particle/Mesh method
• “Particles” connected into materials
Pow3/SP 256^3
Pow3/SP 512^3
P4/Myr 512^2x256
Time per timestep
time (secs)
100
80
60
40
Code Size in Lines
20
0
1
2
4
8
# procs
16
32
64
12
8
Joint work with Ed Givelberg, Armando Solar-Lezama
Fortran
Titanium
8000
4000
Kathy Yelick, 9
The 3 P’s of Parallel Computing
• Productivity
• Global address space supports complex shared structures
• High level constructs simplify programming
• Performance
• PGAS Languages are Faster than two-sided MPI
• Better match to most HPC networks
• Some surprising hints on performance tuning
• Send early and often is sometimes best
• Portability
• These languages are nearly ubiquitous
Kathy Yelick, 10
PGAS Languages: High Performance
Strategy for acceptance of a new language
• Make it run faster than anything else
Keys to high performance
• Parallelism:
• Scaling the number of processors
• Maximize single node performance
• Generate friendly code or use tuned libraries (BLAS, FFTW, etc.)
• Avoid (unnecessary) communication cost
• Latency, bandwidth, overhead
• Berkeley UPC and Titanium use GASNet communication layer
• Avoid unnecessary delays due to dependencies
• Load balance; Pipeline algorithmic dependencies
Kathy Yelick, 11
One-Sided vs Two-Sided
one-sided put message
address
data payload
two-sided message
message id
host
CPU
network
interface
data payload
memory
• A one-sided put/get message can be handled directly by a network
interface with RDMA support
• Avoid interrupting the CPU or storing data from CPU (preposts)
• A two-sided messages needs to be matched with a receive to
identify memory address to put data
• Offloaded to Network Interface in networks like Quadrics
• Need to download match tables to interface (from host)
Kathy Yelick, 12
Performance Advantage of One-Sided
Communication: GASNet vs MPI
900
GASNet put (nonblock)"
MPI Flood
700
Bandwidth (MB/s)
(up is good)
800
600
500
Relative BW (GASNet/MPI)
400
2. 4
2. 2
300
2. 0
1. 8
200
1. 6
1. 4
1. 2
100
1. 0
10
1000 S i z e ( b y t e s )100000
10000000
0
10
100
1,000
10,000
100,000
1,000,000
10,000,000
Size (bytes)
• Opteron/InfiniBand (Jacquard at NERSC):
• GASNet’s vapi-conduit and OSU MPI 0.9.5 MVAPICH
• Half power point (N ½ ) differs by one order of magnitude
Joint work with Paul Hargrove and Dan Bonachea
Kathy Yelick, 13
GASNet: Portability and High-Performance
8-byte Roundtrip Latency
24.2
25
22.1
MPI ping-pong
Roundtrip Latency (usec)
(down is good)
20
GASNet put+sync
18.5
17.8
15
14.6
13.5
9.6
10
9.5
8.3
6.6
6.6
4.5
5
0
Elan3/Alpha
Elan4/IA64
Myrinet/x86
IB/G5
IB/Opteron
SP/Fed
GASNet better for latency across machines
Joint work with UPC Group; GASNet design by Dan Bonachea
Kathy Yelick, 14
GASNet: Portability and High-Performance
Flood Bandwidth for 2MB messages
Percent HW peak (BW in MB)
(up is good)
100%
90%
857
244
858
225
228
799
795
255
1504
1490
80%
610
70%
630
60%
50%
40%
30%
20%
10%
MPI
GASNet
0%
Elan3/Alpha
Elan4/IA64
Myrinet/x86
IB/G5
IB/Opteron
SP/Fed
GASNet at least as high (comparable) for large messages
Joint work with UPC Group; GASNet design by Dan Bonachea
Kathy Yelick, 15
GASNet: Portability and High-Performance
Flood Bandwidth for 4KB messages
100%
223
90%
231
Percent HW peak
(up is good)
80%
70%
MPI
763
714
702
GASNet
679
190
152
60%
420
50%
40%
750
547
252
30%
20%
10%
0%
Elan3/Alpha
Elan4/IA64
Myrinet/x86
IB/G5
IB/Opteron
SP/Fed
GASNet excels at mid-range sizes: important for overlap
Joint work with UPC Group; GASNet design by Dan Bonachea
Kathy Yelick, 16
Case Study 2: NAS FT
• Performance of Exchange (Alltoall) is critical
• 1D FFTs in each dimension, 3 phases
• Transpose after first 2 for locality
• Bisection bandwidth-limited
• Problem as #procs grows
• Three approaches:
• Exchange:
• wait for 2nd dim FFTs to finish, send 1
message per processor pair
• Slab:
• wait for chunk of rows destined for 1
proc, send when ready
• Pencil:
• send each row as it completes
Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea
Kathy Yelick, 17
Overlapping Communication
• Goal: make use of “all the wires all the time”
• Schedule communication to avoid network backup
• Trade-off: overhead vs. overlap
• Exchange has fewest messages, less message overhead
• Slabs and pencils have more overlap; pencils the most
• Example: Class D problem on 256 Processors
Exchange (all data at once)
512 Kbytes
Slabs (contiguous rows that go to 1
processor)
64 Kbytes
Pencils (single row)
16 Kbytes
Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea
Kathy Yelick, 18
NAS FT Variants Performance Summary
Best MFlop rates for all NAS FT Benchmark versions
1000
.5 Tflops
Best NAS Fortran/MPI
Best MPI
Best UPC
MFlops per Thread
800
600
400
200
0
56
et 6 4
nd 2
a
B
i
Myr in
Infin
3 256
Elan
3 512
Elan
4 256
Elan
4 512
Elan
• Slab is always best for MPI; small message cost too high
• Pencil is always best for UPC; more overlap
Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea
Kathy Yelick, 19
Case Study 2: LU Factorization
• Direct methods have complicated dependencies
• Especially with pivoting (unpredictable communication)
• Especially for sparse matrices (dependence graph with
holes)
• LU Factorization in UPC
• Use overlap ideas and multithreading to mask latency
• Multithreaded: UPC threads + user threads + threaded BLAS
• Panel factorization: Including pivoting
• Update to a block of U
• Trailing submatrix updates
• Status:
• Dense LU done: HPL-compliant
Joint work with
Parry Husbands
• Sparse
version
underway
Kathy Yelick, 20
UPC HPL Performance
X1 Linpack Performance
Opteron Cluster
Linpack
Performance
1400
Altix Linpack
Performance
160
MPI/HPL
1200
UPC
140
200
120
800
100
600
100
400
GFlop/s
150
GFlop/s
GFlop/s
1000
MPI/HPL
80
60
UPC
40
MPI/HPL
UPC
•MPI HPL numbers
from HPCC
database
•Large scaling:
• 2.2 TFlops on 512p,
• 4.4 TFlops on 1024p
(Thunder)
50
200
20
0
0
0
60
X1/64
X1/128
Opt/64
Alt/32
• Comparison to ScaLAPACK on an Altix, a 2 x 4 process grid
• ScaLAPACK (block size 64) 25.25 GFlop/s (tried several block sizes)
• UPC LU (block size 256) - 33.60 GFlop/s, (block size 64) - 26.47 GFlop/s
• n = 32000 on a 4x4 process grid
• ScaLAPACK - 43.34 GFlop/s (block size = 64)
• UPC - 70.26 Gflop/s (block size = 200)
Kathy Yelick, 21
Joint work with Parry Husbands
The 3 P’s of Parallel Computing
• Productivity
• Global address space supports complex shared structures
• High level constructs simplify programming
• Performance
• PGAS Languages are Faster than two-sided MPI
• Some surprising hints on performance tuning
• Portability
• These languages are nearly ubiquitous
• Source-to-source translators are key
• Combined with portable communication layer
• Specialized compilers are useful in some cases
Kathy Yelick, 22
Portability of Titanium and UPC
• Titanium and the Berkeley UPC translator use a similar model
• Source-to-source translator (generate ISO C)
• Runtime layer implements global pointers, etc
• Common communication layer (GASNet)
Also used by gcc/upc
• Both run on most PCs, SMPs, clusters & supercomputers
• Support Operating Systems:
• Linux, FreeBSD, Tru64, AIX, IRIX, HPUX, Solaris, Cygwin, MacOSX, Unicos, SuperUX
• UPC translator somewhat less portable: we provide a http-based compile server
• Supported CPUs:
• x86, Itanium, Alpha, Sparc, PowerPC, PA-RISC, Opteron
• GASNet communication:
• Myrinet GM, Quadrics Elan, Mellanox Infiniband VAPI, IBM LAPI, Cray X1, SGI Altix,
Cray/SGI SHMEM, and (for portability) MPI and UDP
• Specific supercomputer platforms:
• HP AlphaServer, Cray X1, IBM SP, NEC SX-6, Cluster X (Big Mac), SGI Altix 3000
• Underway: Cray XT3, BG/L (both run over MPI)
• Can be mixed with MPI, C/C++, Fortran
Joint work with Titanium and UPC groups
Kathy Yelick, 23
Portability of PGAS Languages
Other compilers also exist for PGAS Languages
• UPC
•
•
•
•
Gcc/UPC by Intrepid: runs on GASNet
HP UPC for AlphaServers, clusters, …
MTU UPC uses HP compiler on MPI (source to source)
Cray UPC
• Co-Array Fortran:
• Cray CAF Compiler: X1, X1E
• Rice CAF Compiler (on ARMCI or GASNet), John Mellor-Crummey
•
•
•
•
Source to source
Processors: Pentium, Itanium2, Alpha, MIPS
Networks: Myrinet, Quadrics, Altix, Origin, Ethernet
OS: Linux32 RedHat, IRIS, Tru64
NB: source-to-source requires cooperation by backend compilers
Kathy Yelick, 24
Summary
• PGAS languages offer performance advantages
• Good match to RDMA support in networks
• Smaller messages may be faster:
• make better use of network: postpone bisection bandwidth pain
• can also prevent cache thrashing for packing
• PGAS languages offer productivity advantage
• Order of magnitude in line counts for grid-based code in Titanium
• Push decisions about packing/not into runtime for portability
(advantage of language with translator vs. library approach)
• Source-to-source translation
• The way to ubiquity
• Complement highly tuned machine-specific compilers
Kathy Yelick, 25
End of Slides
PGAS Languages
26
Kathy Yelick
Productizing BUPC
• Recent Berkeley UPC release
• Support full 1.2 language spec
• Supports collectives (tuning ongoing); memory model compliance
• Supports UPC I/O (naïve reference implementation)
• Large effort in quality assurance and robustness
• Test suite: 600+ tests run nightly on 20+ platform configs
• Tests correct compilation & execution of UPC and GASNet
• >30,000 UPC compilations and >20,000 UPC test runs per night
• Online reporting of results & hookup with bug database
• Test suite infrastructure extended to support any UPC compiler
• now running nightly with GCC/UPC + UPCR
• also support HP-UPC, Cray UPC, …
• Online bug reporting database
• Over >1100 reports since Jan 03
• > 90% fixed (excl. enhancement requests)
Kathy Yelick, 27
Benchmarking
• Next few UPC and MPI application benchmarks
use the following systems
•
•
•
•
Myrinet: Myrinet 2000 PCI64B, P4-Xeon 2.2GHz
InfiniBand: IB Mellanox Cougar 4X HCA, Opteron 2.2GHz
Elan3: Quadrics QsNet1, Alpha 1GHz
Elan4: Quadrics QsNet2, Itanium2 1.4GHz
Kathy Yelick, 30
PGAS Languages: Key to High Performance
One way to gain acceptance of a new language
• Make it run faster than anything else
Keys to high performance
• Parallelism:
• Scaling the number of processors
• Maximize single node performance
• Generate friendly code or use tuned libraries (BLAS, FFTW, etc.)
• Avoid (unnecessary) communication cost
• Latency, bandwidth, overhead
• Avoid unnecessary delays due to dependencies
• Load balance
• Pipeline algorithmic dependencies
Kathy Yelick, 31
Hardware Latency
• Network latency is not expected to improve significantly
• Overlapping communication automatically (Chen)
• Overlapping manually in the UPC applications (Husbands, Welcome,
Bell, Nishtala)
• Language support for overlap (Bonachea)
Kathy Yelick, 32
Effective Latency
Communication wait time from other factors
• Algorithmic dependencies
• Use finer-grained parallelism, pipeline tasks (Husbands)
• Communication bandwidth bottleneck
• Message time is: Latency + 1/Bandwidth * Size
• Too much aggregation hurts: wait for bandwidth term
• De-aggregation optimization: automatic (Iancu);
• Bisection bandwidth bottlenecks
• Spread communication throughout the computation (Bell)
Kathy Yelick, 33
Fine-grained UPC vs. Bulk-Synch MPI
• How to waste money on supercomputers
• Pack all communication into single message (spend
memory bandwidth)
• Save all communication until the last one is ready (add
effective latency)
• Send all at once (spend bisection bandwidth)
• Or, to use what you have efficiently:
• Avoid long wait times: send early and often
• Use “all the wires, all the time”
• This requires having low overhead!
Kathy Yelick, 34
What You Won’t Hear Much About
• Compiler/runtime/gasnet bug fixes, performance
tuning, testing, …
• >13,000 e-mail messages regarding cvs checkins
• Nightly regression testing
• 25 platforms, 3 compilers (head, opt-branch, gcc-upc),
• Bug reporting
• 1177 bug reports, 1027 fixed
• Release scheduled for later this summer
• Beta is available
• Process significantly streamlined
Kathy Yelick, 35
Take-Home Messages
• Titanium offers tremendous gains in productivity
• High level domain-specific array abstractions
• Titanium is being used for real applications
• Not just toy problems
• Titanium and UPC are both highly portable
• Run on essentially any machine
• Rigorously tested and supported
• PGAS Languages are Faster than two-sided MPI
• Better match to most HPC networks
• Berkeley UPC and Titanium benchmarks
• Designed from scratch with one-side PGAS model
• Focus on 2 scalability challenges: AMR and Sparse LU
Kathy Yelick, 36
Titanium Background
• Based on Java, a cleaner C++
• Classes, automatic memory management, etc.
• Compiled to C and then machine code, no JVM
• Same parallelism model at UPC and CAF
• SPMD parallelism
• Dynamic Java threads are not supported
• Optimizing compiler
• Analyzes global synchronization
• Optimizes pointers, communication, memory
Kathy Yelick, 37
Do these Features Yield Productivity?
MG Line Count Comparison
CG Line Count Comparison
Computation
Communication
Declarations
203
700
600
500
400
552
300
200
92
100
0
60
58
Fortran w/ MPI
Titanium
36
Productive Lines of Code
160
Computation
Communication
Declarations
800
140
120
100
99
80
44
60
28
40
27
20
35
14
0
Fortran w/ MPI
Titanium
Language
Language
CG Class D Speedup - G5/IB
Speedup (Over Best 64
Proc Case)
MG Class D Speedup - Opteron/IB
Speedup (Over Best 32
Proc Case)
Productive Lines of Code
900
140
Linear Speedup
Fortran w/MPI
Titanium
120
100
80
60
40
20
0
0
50
100
150
Processors
Joint work with Kaushik Datta, Dan Bonachea
400
350
300
250
200
150
100
50
0
Linear Speedup
Fortran w/MPI
Titanium
0
50
100
150
200
250
300
Processors
Kathy Yelick, 38
GASNet/X1 Performance
single word get
14
13
Shm em
GASNet
12
MPI
14
13
Get Lat ency (m icroseconds)
Put per m essage gap (m icroseconds)
single word put
11
10
9
8
7
6
5
4
3
2
1
12
Shm em
GASNet
11
10
9
8
7
6
5
4
3
2
1
0
0
1
2
RMW
4
Sc a la r
8
16
32
64
128
Ve c t or
256
512 1024 2048
b c op y()
1
RMW
Message Size (byt es)
2
4
Sc a la r
8
16
32
64
Ve c t or
128
256 512 1024 2048
b c op y()
Message Size (byt es)
• GASNet/X1 improves small message performance over shmem and MPI
• Leverages global pointers on X1
• Highlights advantage of languages vs. library approach
Joint work with Christian Bell, Wei Chen and Dan Bonachea
Kathy Yelick, 39
High Level Optimizations in Titanium
• Irregular communication can be expensive
• “Best” strategy differs by data size/distribution and machine
parameters
• E.g., packing, sending bounding boxes, fine-grained are
• Use of runtime
optimizations
Itanium/Myrinet Speedup Comparison
• Inspector-executor
Average and maximum speedup
of the Titanium version relative to
the Aztec version on 1 to 16
processors
Joint work with Jimmy Su
1.5
speedup
• Performance on
Sparse MatVec Mult
• Results: best strategy
differs within the
machine on a single
matrix (~ 50% better)
Speedup relative to MPI code (Aztec library)
1.6
1.4
1.3
1.2
1.1
1
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22
matrix number
average speedup
maximum speedup
Kathy Yelick, 40
Source to Source Strategy
• Source-to-source translation strategy
• Tremendous portability advantage
• Still can perform significant optimizations
• Use of “restrict” pointers in C
• Understand Multi-D array
indexing (Intel/Itanium issue)
• Support for pragmas like
IVDEP
• Robust vectorizators (X1,
SSE, NEC,…)
Performance Ratio C / UPC
• Relies on high quality back-end compilers and some
coaxing in code generation
Livermore Loops on Cray X1 (single Node)
48x
4
3.5
3
2.5
2
1.5
1
0.5
0
1 2 3 4 5
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
• On machines with integrated shared memory
hardware need access to shared memory operations
Joint work with Jimmy Su
Kathy Yelick, 41