forked from jedbrown/talks
-
Notifications
You must be signed in to change notification settings - Fork 0
/
20130730-GPUSMP.tex
423 lines (371 loc) · 14.5 KB
/
20130730-GPUSMP.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
% \documentclass[handout]{beamer}
\documentclass{beamer}
\mode<presentation>
{
\usetheme{default}
\usefonttheme[onlymath]{serif}
% \usetheme{Singapore}
% \usetheme{Warsaw}
% \usetheme{Malmoe}
% \useinnertheme{circles}
% \useoutertheme{infolines}
% \useinnertheme{rounded}
\setbeamercovered{transparent=20}
}
\usepackage[english]{babel}
\usepackage[latin1]{inputenc}
\usepackage{alltt,listings,multirow,ulem,siunitx}
\usepackage[absolute,overlay]{textpos}
\TPGrid{1}{1}
\usepackage{pdfpages}
\usepackage{multimedia}
\usepackage{multicol}
\newcommand\hmmax{0}
\newcommand\bmmax{0}
\usepackage{bm}
\usepackage{comment}
% font definitions, try \usepackage{ae} instead of the following
% three lines if you don't like this look
\usepackage{mathptmx}
\usepackage[scaled=.90]{helvet}
% \usepackage{courier}
\usepackage[T1]{fontenc}
\usepackage{tikz}
\usetikzlibrary{decorations.pathreplacing}
\usetikzlibrary{shadows,arrows,shapes.misc,shapes.arrows,shapes.multipart,arrows,decorations.pathmorphing,backgrounds,positioning,fit,petri,calc,shadows,chains,matrix}
% \usepackage{pgfpages}
% \pgfpagesuselayout{4 on 1}[a4paper,landscape,border shrink=5mm]
\usepackage{JedMacros}
\newcommand{\timeR}{t_{\mathrm{R}}}
\newcommand{\timeW}{t_{\mathrm{W}}}
\newcommand{\mglevel}{\ensuremath{\ell}}
\newcommand{\mglevelcp}{\ensuremath{\mglevel_{\mathrm{cp}}}}
\newcommand{\mglevelcoarse}{\ensuremath{\mglevel_{\mathrm{coarse}}}}
\newcommand{\mglevelfine}{\ensuremath{\mglevel_{\mathrm{fine}}}}
%solution and residual
\newcommand{\vx}{\ensuremath{x}}
\newcommand{\vc}{\ensuremath{\hat{x}}}
\newcommand{\vr}{\ensuremath{r}}
\newcommand{\vb}{\ensuremath{b}}
%operators
\newcommand{\vA}{\ensuremath{A}}
\newcommand{\vP}{\ensuremath{I_H^h}}
\newcommand{\vS}{\ensuremath{S}}
\newcommand{\vR}{\ensuremath{I_h^H}}
\newcommand{\vI}{\ensuremath{\hat I_h^H}}
\newcommand{\vV}{\ensuremath{\mathbf{V}}}
\newcommand{\vF}{\ensuremath{F}}
\newcommand{\vtau}{\ensuremath{\mathbf{\tau}}}
\title{GPU-accelerated smoothed aggregation algebraic multigrid: Multi-node scalability and versatility}
\author{{\bf Jed Brown}\inst{1} and Steven Dalton\inst{2} \\
{\small thanks to Karl Rupp\inst{1} and Luke Olson\inst{2}}}
% - Use the \inst command only if there are several affiliations.
% - Keep it simple, no one is interested in your street address.
\institute
{
\inst{1}{Mathematics and Computer Science Division, Argonne National Laboratory} \\
\inst{2}{Department of Computer Science, University of Illinois at Urbana-Champaign}
}
\date{GPU-SMP, Changchun, 2013-07-30}
% This is only inserted into the PDF information catalog. Can be left
% out.
\subject{Talks}
% If you have a file called "university-logo-filename.xxx", where xxx
% is a graphic format that can be processed by latex or pdflatex,
% resp., then you can add a logo as follows:
% \pgfdeclareimage[height=0.5cm]{university-logo}{university-logo-filename}
% \logo{\pgfuseimage{university-logo}}
% Delete this, if you do not want the table of contents to pop up at
% the beginning of each subsection:
% \AtBeginSubsection[]
% {
% \begin{frame}<beamer>
% \frametitle{Outline}
% \tableofcontents[currentsection,currentsubsection]
% \end{frame}
% }
\AtBeginSection[]
{
\begin{frame}<beamer>
\frametitle{Outline}
\tableofcontents[currentsection]
\end{frame}
}
% If you wish to uncover everything in a step-wise fashion, uncomment
% the following command:
% \beamerdefaultoverlayspecification{<+->}
\begin{document}
\lstset{language=C}
\normalem
\begin{frame}
\titlepage
\end{frame}
\section{Smoothed aggregation and fine-grained parallelism}
\begin{frame}[fragile]{Smoothed aggregation}
\begin{itemize}
\item Rapid-coarsening algebraic multigrid/domain decomposition
\item Robust for vector-valued problems (elasticity, near-null space)
\item Conventional algorithmic structure given fine-grid operator $A$:
\begin{enumerate}
\item Create strength-of-connection graph $G$
\item Dropping small edges to create thresholded graph $\tilde G$ (important for anisotropy and $p>1$)
\item Compute aggregates, usually via $MIS(k)$ for root nodes, $k=2$
\item Tentative interpolation $T$: columns are coarse basis functions on aggregates
\item Estimate spectrum of $A$ and define smoother $S = 1 - \omega D^{-1} A$
\item \alert<2>{Compute smoothed prolongator $P = S T$}
\item \alert<2>{Compute coarse operator via Galerkin product $A_c = P^T A P$}
\end{enumerate}
\item Cycling: $V$-cycle, $F$-cycle, $W$-cycle
\begin{figure}
\centering
\begin{tikzpicture}
[>=stealth,
every node/.style={inner sep=2pt},
restrict/.style={thick},
prolong/.style={thick},
mglevel/.style={rounded rectangle,draw=blue!50!black,fill=blue!20,thick,minimum size=4mm},
]
\begin{scope}\scriptsize
\newcommand\mgdx{4.0em}
\newcommand\mgdy{3.0em}
\newcommand\mgl[1]{(pow(2,#1+1))}
\newcommand\mgloc[4]{(#1 + #4*\mgdx*#3,#2 + \mgdy*#3)}
\node[mglevel] (down0) at \mgloc{0}{0}{2}{-1} {\mglevel$_{fine}$};
\node[mglevel] (down1) at \mgloc{0}{0}{1}{-1} {};
\node[mglevel] (coarse) at \mgloc{0}{0}{0}{-1} {\mglevel$_{coarse}$};
\node[mglevel] (up1) at \mgloc{0}{0}{1}{1} {};
\node[mglevel] (up0) at \mgloc{0}{0}{2}{1} {\mglevel$_{fine}$};
\path[->,restrict] (down0) edge node [above right] {$b_c = P^T r$} (down1)
(down1) edge node [above right] {$b_c = P^T r$} (coarse);
\path[->,prolong] (coarse) edge node [right] {$P \hat x$} (up1)
(up1) edge node [right] {$P \hat x$} (up0);
% grids
\newcommand\mghx{0.9*\mgdx}
\newcommand\mghy{0.9*\mgdy}
\draw[shift=\mgloc{-5*\mgdx}{0}{2}{0},
xstep=\mghy/\mgl{2},
ystep=\mghy/\mgl{2}]
(-0.5*\mghy,-0.5*\mghy) grid (0.5*\mghy,0.5*\mghy);
\draw[shift=\mgloc{-5*\mgdx}{0}{1}{0},
xstep=\mghy/\mgl{1},
ystep=\mghy/\mgl{1}]
(-0.5*\mghy,-0.5*\mghy) grid (0.5*\mghy,0.5*\mghy);
\draw[shift=\mgloc{-5*\mgdx}{0}{0}{0},
xstep=\mghy/\mgl{0},
ystep=\mghy/\mgl{0}]
(-0.5*\mghy,-0.5*\mghy) grid (0.5*\mghy,0.5*\mghy);
\end{scope}
\end{tikzpicture}
\label{fig:MG}
\end{figure}
\end{itemize}
\end{frame}
\begin{frame}{Aggregation and coarsening}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/MG/LubyAgg1} \qquad \qquad
\includegraphics[width=0.4\textwidth]{figures/MG/LubyAgg2}
\caption{$MIS(2)$ finds root nodes (red), defines aggregates, contraction $T$.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figures/MG/CoarseUnsmoothedRoot} \qquad \qquad
\includegraphics[width=0.4\textwidth]{figures/MG/CoarseSmoothedRoot}
\caption{Coarse graph: unsmoothed $T^T A T$, smoothed $P^T A P = (ST)^T A (ST)$.}
\end{figure}
\end{frame}
\begin{frame}{Fine-grained parallelism for setup}
\begin{itemize}
\item SpMM: Sparse matrix-matrix products
\begin{enumerate}
\item $P = S T = (1 - \omega D^{-1}A) T$
\item $A P$
\item $P^T (A P) = R (A P)$
\end{enumerate}
\item Three GPU phases to SpMM: Expand, Sort, Contract
\item Pros of ESC
\begin{itemize}
\item Low memory usage per thread
\item High-level primitives
\item Predictable performance
\end{itemize}
\item Cons of ESC
\begin{itemize}
\item Large global memory usage, can be hard to predict
\item High bandwidth to global memory
\item Multiplications and additions in separate phases---no FMA
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Fine-grained parallelism in SpMM}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figures/MG/SACUSPExpand}
\end{figure}
\begin{itemize}
\item Enumerate all scalar products contributing to row of product, $\hat C$
\item Implemented using \texttt{scan} and \texttt{gather}
\item Radix sort contributions to each row (two calls to \texttt{sort})
\item Contract row: \texttt{reduce\_by\_key}
\end{itemize}
\end{frame}
\begin{frame}{Sorting optimization and performance}
\begin{itemize}
\item Sorting optimization, similar to B40C (Merrill and Grimshaw, 2011)
\begin{itemize}
\item Bin expanded rows ($\hat C$) by length
\item Short rows: single thread via optimal sorting network
\item Medium-length: radix sort within block (shared memory)
\item Long rows: radix sort in global memory
\end{itemize}
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figures/MG/SACUSPMaticesFE.png}
\caption{Time to compute $AP$ (ms) on Tesla C2075 and {\bf speedup}.}
\end{figure}
\end{frame}
\begin{frame}{CUSP Performance summary}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figures/MG/SACUSPSpeedupAP}
\end{figure}
\begin{itemize}
\item New CUSP SpMM is faster than CUSPARSE for all test matrices.
\item Sorting optimization faster except for very irregular graph.
\end{itemize}
\end{frame}
\begin{frame}{Memory overhead from expansion}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/MG/SACUSPExpansionFactor} \quad
\includegraphics[width=0.48\textwidth]{figures/MG/SACUSPContractionFactor}
\caption{Scalar Poisson: Expansion factor $nnz(\hat C)/nnz(A)$, contraction $nnz(\hat C)/nnz(C)$}
\end{figure}
\vspace{-2em}
\begin{itemize}
\item 3D has much higher variability by row
\item For elasticity, expansion factor is larger by 3x (for 3D)
\item Implementation could batch to limit total memory usage
\end{itemize}
\end{frame}
\begin{frame}{Memory Streaming Efficiency (GB/Joule)}
\begin{figure}
\centering
\includegraphics<1>[width=0.9\textwidth]{figures/hardware/mem-bandwidth-per-watt}
\includegraphics<2->[width=0.9\textwidth]{figures/hardware/mem-bandwidth-per-watt-with-host}
\end{figure}
\begin{itemize}
\item<2-> Include 100 W host CPU paired with each GPU device
\item<3> Implications for SpMM
\begin{itemize}
\item Expand-Sort-Compress: multiple accesses to expanded representation
\item Compare to CPU: no expanded representation, only accessed once
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Distributed memory SpMM}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{figures/Mat/parallelSparseMatrix}
\end{figure}
\begin{itemize}
\item Matrices $A$ and $P$ partitioned by rows
\item Stored as $P_{\text{rank}} = [\hat P_d, \hat P_o]_{\text{rank}}$, map $\hat p$: local $\hat P_o$ to global col
\item CPU: unpack off-process rows $\check P_{\text{rank}} = [\check P_d, \check P_o]_{\text{rank}}$
\begin{equation*}
\begin{bmatrix}
AP = \widehat{AP}_d & \widehat{AP}_o
\end{bmatrix}
=
\begin{bmatrix}
\hat A_d & \hat A_o
\end{bmatrix}
\begin{bmatrix}
\hat P_d & \alert<2>{\hat P_o} \\
\check P_d & \alert<2>{\check P_o}
\end{bmatrix}
\end{equation*}
\item<2-> \alert<2>{$\hat P_o$ and $\check P_o$ have different column spaces (defined by $\hat p$ and $\check p$).}
\item<3> GPU: Global column okay with GPU (already sorting); all of $\check P$ must be transferred
\end{itemize}
\end{frame}
\section{Versatility}
\begin{frame}{Versatility of accelerators for multiphysics engineering/science}
\begin{itemize}
\item Library implementations for different sizes
\item Do not want to recompile for every model variant
\begin{itemize}
\item Provenance/reproducibility, testing, build system complexity
\item Many variants used simultaneously in same application
\item Model changes in optimization loop, large state to carry to next iteration
\end{itemize}
\item Material and chemistry libraries: high variability, multiphysics
\begin{itemize}
\item 5-3000 chemical species, crystal grains, energy groups, etc
\item 5-10000 flops for equation of state (humidity, combustion, ionization)
\begin{itemize}
\item Defined implicitly: need local Newton iteration to evaluate (convergence test)
\end{itemize}
\item Modified constantly by domain scientists (not software engineers)
\end{itemize}
\item Do not need optimality in all cases, should degrade \emph{gracefully}
\begin{itemize}
\item Warp size, number of registers, shared memory size
\item Data dependencies/convergence irregularity at quadrature points
\item Memory reuse conflicts with vectorization
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Example: Climate}
\begin{itemize}
\item Many chemical species: $\sim 100$ atmosphere, $\sim 20$ ocean
\item Users change reaction mechanisms and selected tracers
\item Tridiagonal solves pervasive due to stiffness in vertical (logarithmic depth)
\item Complicated equations of state: 32-term rational powers and exponentials
\item Constant turn-around time desired as resolution is increased
\begin{itemize}
\item CFL condition: $\Delta t \sim \Delta x$
\item 2x resolution requires steps/second to double
\item Already need 1000 steps/second
\end{itemize}
\item Different code components maintained by different science teams
\end{itemize}
\end{frame}
\begin{frame}{Where we are now: $QR$ factorization with MKL on MIC}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figures/hardware/MKL-dgeqrf-MIC}
\end{figure}
\begin{itemize}
\item Figure compares two CPU sockets (230W TDP) to one MIC (300W TDP plus host)
\item Performance/Watt only breaks even at largest problem sizes
\item $10^4 \times 10^4$ matrix takes 667 GFlops: about 2 seconds
\item MIC cannot strong scale, no more energy efficient/cost effective
\end{itemize}
\end{frame}
\begin{frame}{Where are we now: latencies (c/o Karl Rupp)}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figures/hardware/karlrupp-vector-timings-7}
\end{figure}
\vspace{-1em}
\begin{itemize}
\item GPU: kernel launch overhead of 10--20 $\mu$s
\item OpenCL toolchain for MIC is atrocious
\item MIC latencies with OpenMP are decent
\end{itemize}
\end{frame}
\begin{frame}{\texttt{MPI\_Allreduce} performance, c/o Paul Fischer}
\includegraphics[width=\textwidth]{figures/hardware/FischerBGQAllReduce.png}
\end{frame}
\begin{frame}\LARGE
\begin{itemize}
\item Maximize science per Watt
\item Versatility is critical for engineering and science
\item Huge scope remains at problem formulation
\item Raise level of abstraction at which a problem is formally specified
\item Algorithmic optimality is crucial
\end{itemize}
\end{frame}
\end{document}