Title: Minimizing Communication in Linear Algebra
1Minimizing Communication in Linear Algebra
2Outline
- What is communication and why is it important
to avoid? - Goal Algorithms that communicate as little as
possible for - BLAS, Axb, QR, eig, SVD, whole programs,
- Dense or sparse matrices
- Sequential or parallel computers
- Direct methods (BLAS, LU, QR, SVD, other
decompositions) - Communication lower bounds for all these problems
- Algorithms that attain them (all dense linear
algebra, some sparse) - Mostly not in LAPACK or ScaLAPACK (yet)
- Iterative methods Krylov subspace methods for
Axb, Ax?x - Communication lower bounds, and algorithms that
attain them (depending on sparsity structure) - Not in any libraries (yet)
- Just highlights
- Preliminary speed-up results
- Up to 8x measured (or 8x), 81x modeled
3Collaborators
- Jim Demmel, UCB EECS
- Ioana Dumitriu, U. Washington
- Laura Grigori, INRIA
- Ming Gu, UCB Math
- Mark Hoemmen, UCB EECS
- Olga Holtz, UCB Math TU Berlin
- Julien Langou, U. Colorado Denver
- Marghoob Mohiyuddin, UCB EECS
- Oded Schwartz , TU Berlin
- Hua Xiang, INRIA
- Kathy Yelick, UCB EECS NERSC
- BeBOP group, bebop.cs.berkeley.edu
4Motivation (1/2)
- Algorithms have two costs
- Arithmetic (FLOPS)
- Communication moving data between
- levels of a memory hierarchy (sequential case)
- processors over a network (parallel case).
5Motivation (2/2)
- Running time of an algorithm is sum of 3 terms
- flops time_per_flop
- words moved / bandwidth
- messages latency
- Time_per_flop ltlt 1/ bandwidth ltlt latency
- Gaps growing exponentially with time
Annual improvements Annual improvements Annual improvements Annual improvements
Time_per_flop Bandwidth Latency
Network 26 15
DRAM 23 5
59
6Direct linear algebra Prior Work on Matmul
- Assume n3 algorithm (i.e. not Strassen-like)
- Sequential case, with fast memory of size M
- Lower bound on words moved to/from slow memory
? (n3 / M1/2 ) Hong, Kung, 81 - Attained using blocked algorithms
- Parallel case on P processors
- Let NNZ be total memory needed assume load
balanced - Lower bound on words communicated
? (n3 /(P NNZ )1/2 )
Irony, Tiskin, Toledo, 04
NNZ (name of alg) Lower bound on words Attained by
3n2 (2D alg) ? (n2 / P1/2 ) Cannon, 69
3n2 P1/3 (3D alg) ? (n2 / P2/3 ) Johnson,93
7Lower bound for all direct linear algebra
- Let M fast memory size per processor
- words_moved by at least one processor
- (flops / M1/2 )
- messages_sent by at least one processor
- (flops / M3/2 )
- Holds for
- BLAS, LU, QR, eig, SVD,
- Some whole programs (sequences of these
operations, no matter how they are interleaved,
eg computing Ak) - Dense and sparse matrices (where flops ltlt n3 )
- Sequential and parallel algorithms
- Some graph-theoretic algorithms (eg
Floyd-Warshall) - See BDHS09
8Can we attain these lower bounds?
- Do conventional dense algorithms as implemented
in LAPACK and ScaLAPACK attain these bounds? - Mostly not
- If not, are there other algorithms that do?
- Yes
- Goals for algorithms
- Minimize words ? (flops/ M1/2 )
- Minimize messages ? (flops/ M3/2 )
- Minimize for multiple memory hierarchy levels
- Recursive, Cache-oblivious algorithms would be
simplest - Fewest flops when matrix fits in fastest memory
- Cache-oblivious algorithms dont always attain
this - Only a few sparse algorithms so far
9Summary of dense sequential algorithms attaining
communication lower bounds
- Algorithms shown minimizing Messages use
(recursive) block layout - Not possible with columnwise or rowwise layouts
- Many references (see reports), only some shown,
plus ours - Cache-oblivious are underlined, Green are ours,
? is unknown/future work
Algorithm 2 Levels of Memory 2 Levels of Memory Multiple Levels of Memory Multiple Levels of Memory
Words Moved and Messages Words Moved and Messages
BLAS-3 Usual blocked or recursive algorithms Usual blocked or recursive algorithms Usual blocked algorithms (nested), or recursive Gustavson,97 Usual blocked algorithms (nested), or recursive Gustavson,97
Cholesky LAPACK (with b M1/2) Gustavson 97 BDHS09 Gustavson,97 Ahmed,Pingali,00 BDHS09 (?same) (?same)
LU with pivoting LAPACK (rarely) Toledo,97 , GDX 08 GDX 08 not partial pivoting Toledo, 97 GDX 08? GDX 08?
QR LAPACK (rarely) Elmroth,Gustavson,98 DGHL08 Frens,Wise,03 but 3x flops DGHL08 Elmroth, Gustavson,98 DGHL08 ? Frens,Wise,03 DGHL08 ?
Eig, SVD Not LAPACK BDD09 randomized, but more flops Not LAPACK BDD09 randomized, but more flops BDD09 BDD09
10Summary of dense 2D parallel algorithms
attaining communication lower bounds
- Assume nxn matrices on P processors, memory
per processor O(n2 / P) - ScaLAPACK assumes best block size b chosen
- Many references (see reports), Green are ours
- Recall lower bounds
- words_moved ?( n2 / P1/2 ) and
messages ?( P1/2 )
Algorithm Reference Factor exceeding lower bound for words_moved Factor exceeding lower bound for messages
Matrix multiply Cannon, 69 1 1
Cholesky ScaLAPACK log P log P
LU GDX08 ScaLAPACK log P log P log P ( N / P1/2 ) log P
QR DGHL08 ScaLAPACK log P log P log3 P ( N / P1/2 ) log P
Sym Eig, SVD BDD09 ScaLAPACK log P log P log3 P N / P1/2
Nonsym Eig BDD09 ScaLAPACK log P P1/2 log P log3 P N log P
11QR of a Tall, Skinny matrix is bottleneckUse
TSQR instead
Q is represented implicitly as a product (tree of
factors) DGHL08
12Minimizing Communication in TSQR
Multicore / Multisocket / Multirack / Multisite /
Out-of-core ?
Can Choose reduction tree dynamically
13Performance of TSQR vs Sca/LAPACK
- Parallel
- Intel Clovertown
- Up to 8x speedup (8 core, dual socket, 10M x 10)
- Pentium III cluster, Dolphin Interconnect, MPICH
- Up to 6.7x speedup (16 procs, 100K x 200)
- BlueGene/L
- Up to 4x speedup (32 procs, 1M x 50)
- Both use Elmroth-Gustavson locally enabled by
TSQR - Sequential
- OOC on PowerPC laptop
- As little as 2x slowdown vs (predicted) infinite
DRAM - LAPACK with virtual memory never finished
- See DGHL08
- General CAQR Communication-Avoiding QR in
progress
14Modeled Speedups of CAQR vs ScaLAPACK
Petascale up to 22.9x IBM Power 5
up to 9.7x Grid up to 11x
Petascale machine with 8192 procs, each at 500
GFlops/s, a bandwidth of 4 GB/s.
15Direct Linear Algebra summary and future work
- Communication lower bounds on words_moved and
messages - BLAS, LU, Cholesky, QR, eig, SVD,
- Some whole programs (compositions of these
operations, no matter how
they are interleaved, eg computing Ak) - Dense and sparse matrices (where flops ltlt n3 )
- Sequential and parallel algorithms
- Some graph-theoretic algorithms (eg
Floyd-Warshall) - Algorithms that attain these lower bounds
- Nearly all dense problems (some open problems)
- A few sparse problems
- Speed ups in theory and practice
- Future work
- Lots to implement, autotune
- Next generation of Sca/LAPACK on heterogeneous
architectures (MAGMA) - Few algorithms in sparse case
- Strassen-like algorithms
- 3D Algorithms (only for Matmul so far), may be
important for scaling
16Avoiding Communication in Iterative Linear Algebra
- k-steps of typical iterative solver for sparse
Axb or Ax?x - Does k SpMVs with starting vector
- Finds best solution among all linear
combinations of these k1 vectors - Many such Krylov Subspace Methods
- Conjugate Gradients, GMRES, Lanczos, Arnoldi,
- Goal minimize communication in Krylov Subspace
Methods - Assume matrix well-partitioned, with modest
surface-to-volume ratio - Parallel implementation
- Conventional O(k log p) messages, because k
calls to SpMV - New O(log p) messages - optimal
- Serial implementation
- Conventional O(k) moves of data from slow to
fast memory - New O(1) moves of data optimal
- Lots of speed up possible (modeled and measured)
- Price some redundant computation
- Much prior work
- See bebop.cs.berkeley.edu
- CG van Rosendale, 83, Chronopoulos and Gear,
89 - GMRES Walker, 88, Joubert and Carey, 92,
Bai et al., 94
17Minimizing Communication of GMRES to solve Axb
- GMRES find x in spanb,Ab,,Akb minimizing
Ax-b 2 - Cost of k steps of standard GMRES vs new GMRES
Standard GMRES for i1 to k w A
v(i-1) MGS(w, v(0),,v(i-1)) update
v(i), H endfor solve LSQ problem with
H Sequential words_moved O(knnz)
from SpMV O(k2n) from MGS Parallel
messages O(k) from SpMV
O(k2 log p) from MGS
Communication-avoiding GMRES W v, Av, A2v,
, Akv Q,R TSQR(W) Tall Skinny
QR Build H from R, solve LSQ
problem Sequential words_moved
O(nnz) from SpMV O(kn) from
TSQR Parallel messages O(1) from
computing W O(log p) from TSQR
- Oops W from power method, precision lost!
18Monomial basis Ax,,Akx fails to converge
A different polynomial basis does converge
19Speed ups of GMRES on 8-core Intel Clovertown
MHDY09
20Communication Avoiding Iterative Linear Algebra
Future Work
- Lots of algorithms to implement, autotune
- Make widely available via OSKI, Trilinos, PETSc,
Python, - Extensions to other Krylov subspace methods
- So far just Lanczos/CG and Arnoldi/GMRES
- BiCGStab, CGS, QMR,
- Add preconditioning
- Block diagonal are easiest
- Hierarchical, semiseparable,
- Works if interactions between distant points can
be compressed
21Summary
221/3
232/3
243/3
25 26Whats new since last time
- Last spoke in this seminar Feb 2009
- Direct case
- Lower bound for QR decomp proven
- Extensions to eigenproblems, SVD, compositions,
- Invented or identified algorithms that attain
lower bounds - Iterative case
- Not presented last time
- Mark Hoemmen has spoken about it
- More good measured speedups (for GMRES TSQR)
27Direct linear algebra Generalizing Prior Work
- Sequential case words moved
- ? (n3 / M1/2 ) ? (flops / (fast_memory_size)1/
2 ) - Parallel case words moved
- ? (n3 /(P NNZ )1/2 ) ? ((n3 /P) / (NNZ/P)1/2
) - ? (flops_per_proc / (memory_per_processor)1/2
) - (true for at least one processor, assuming
balance in either flops or in memory) - In both cases, we let M memory size, and write
words_moved by at least one processor ?
(flops / M1/2 )
28Proof of Communication Lower Bound on C AB
(1/5)
- Proof from Irony/Toledo/Tiskin (2004)
- Original proof, then generalization
- Think of instruction stream being executed
- Looks like add, load, multiply, store,
load, add, - Each load/store moves a word between fast and
slow memory - We want to count the number of loads and stores,
given that we are multiplying n-by-n matrices C
AB using the usual 2n3 flops, possibly reordered
assuming addition is commutative/associative - Assuming that at most M words can be stored in
fast memory - Outline
- Break instruction stream into segments, each with
M loads and stores - Somehow bound the maximum number of flops that
can be done in each segment, call it F - So F segments ? T total flops 2n3 ,
so segments ? T / F - So loads stores M segments ? M T /
F
29Illustrating Segments, for M3
...
30Proof of Communication Lower Bound on C AB
(1/5)
- Proof from Irony/Toledo/Tiskin (2004)
- Original proof, then generalization
- Think of instruction stream being executed
- Looks like add, load, multiply, store,
load, add, - Each load/store moves a word between fast and
slow memory - We want to count the number of loads and stores,
given that we are multiplying n-by-n matrices C
AB using the usual 2n3 flops, possibly reordered
assuming addition is commutative/associative - Assuming that at most M words can be stored in
fast memory - Outline
- Break instruction stream into segments, each with
M loads and stores - Somehow bound the maximum number of flops that
can be done in each segment, call it F - So F segments ? T total flops 2n3 ,
so segments ? T / F - So loads stores M segments ? M T /
F
31Proof of Communication Lower Bound on C AB
(2/5)
- Given segment of instruction stream with M loads
stores, how many adds multiplies (F) can we
do? - At most 2M entries of C, 2M entries of A and/or
2M entries of B can be accessed - Use geometry
- Represent n3 multiplications by n x n x n cube
- One n x n face represents A
- each 1 x 1 subsquare represents one A(i,k)
- One n x n face represents B
- each 1 x 1 subsquare represents one B(k,j)
- One n x n face represents C
- each 1 x 1 subsquare represents one C(i,j)
- Each 1 x 1 x 1 subcube represents one C(i,j)
A(i,k) B(k,j) - May be added directly to C(i,j), or to temporary
accumulator
32Proof of Communication Lower Bound on C AB
(3/5)
k
C face
C(2,3)
C(1,1)
B(3,1)
A(1,3)
j
B(2,1)
A(1,2)
B(1,3)
B(1,1)
A(2,1)
A(1,1)
B face
i
- If we have at most 2M A squares, 2M B
squares, and 2M C squares on faces, how many
cubes can we have?
A face
33Proof of Communication Lower Bound on C AB
(4/5)
x
y
z
y
z
x
(i,k) is in A shadow if (i,j,k) in 3D set
(j,k) is in B shadow if (i,j,k) in 3D set
(i,j) is in C shadow if (i,j,k) in 3D
set Thm (Loomis Whitney, 1949) cubes in
3D set Volume of 3D set (area(A shadow)
area(B shadow) area(C shadow)) 1/2
cubes in black box with side lengths x, y
and z Volume of black box xyz ( xz zy
yx)1/2 (A?s B?s C?s )1/2
34Proof of Communication Lower Bound on C AB
(5/5)
- Consider one segment of instructions with M
loads, stores - Can be at most 2M entries of A, B, C available in
one segment - Volume of set of cubes representing possible
multiply/adds in one segment is (2M 2M
2M)1/2 (2M) 3/2 F - Segments ? 2n3 / F
- Loads Stores M Segments ? M 2n3 / F
n3 / (2M)1/2
- Parallel Case apply reasoning to one processor
out of P - Adds and Muls ? 2n3 / P (at least one proc
does this ) - M n2 / P (each processor gets equal fraction of
matrix) - Load Stores words moved from or to
other procs ? M (2n3 /P) / F M (2n3 /P) /
(2M) 3/2 n2 / (2P)1/2
35How to generalize this lower bound (1/4)
- It doesnt depend on C(i,j) being a matrix entry,
just a unique memory location (same for A(i,k)
and B(k,j) )
- It doesnt depend on C and A not overlapping (or
C and B, or A and B) - Some reorderings may change answer still get a
lower bound
- It doesnt depend on doing n3 multiply/adds
- It doesnt depend on C(i,j) just being a sum of
products
C(i,j) ?k A(i,k)B(k,j)
For all (i,j) ? S Mem(c(i,j))
fij ( gijk ( Mem(a(i,k)) , Mem(b(k,j)) ) for k ?
Sij , some other arguments)
36How to generalize this lower bound (2/4)
- It does assume the presence of an operand
generating a load and/or store how could this
not happen? - Mem(b(k,j)) could be reused in many more gijk
than (P) allows - Ex Compute C(m) A B.m (Matlab notation)
for m1 to t - Can move t1/2 times fewer words than ? (flops /
M1/2 ) - We might generate a result during a segment, use
it, and discard it, with generating any memory
traffic - Turns out QR, eig, SVD all may do this
- Need a different analysis for them later
37How to generalize this lower bound (3/4)
- Need to distinguish Sources, Destinations of
each operand in fast memory during a segment - Possible Sources
- S1 Already in fast memory at start of segment,
or read at most 2M - S2 Created during segment no bound without
more information - Possible Destinations
- D1 Left in fast memory at end of segment, or
written at most 2M - D2 Discarded no bound without more information
- Need to assume no S2/D2 arguments at most 4M of
others
38How to generalize this lower bound (4/4)
- Theorem To evaluate (P) with memory of size M,
where - fij and gijk are nontrivial functions of their
arguments - G is the total number of gijks,
- No S2/D2 arguments
- requires at least G/ (8M)1/2 M slow
memory references
- Simpler words_moved ? (flops / M1/2 )
- Corollary To evaluate (P) requires at least
G/ (81/2 M3/2 ) 1
messages (simpler ? (flops / M3/2 ) ) - Proof maximum message size is M
39Some Corollaries of this lower bound (1/2)
words_moved ? (flops / M1/2 )
?
- Theorem applies to dense or sparse, parallel or
sequential - MatMul, including ATA or A2
- Triangular solve C A-1X
- C(i,j) (X(i,j) - ?klti A(i,k)C(k,j)) / A(i,i)
A lower triangular - C plays double role of b and c in Model (P)
- LU factorization (any pivoting, LU or ILU)
- L(i,j) (A(i,j) - ?kltj L(i,k)U(k,j)) / U(j,j),
U(i,j) A(i,j) - ?klti L(i,k)U(k,j) - L (and U) play double role as c and a (c and b)
in Model (P) - Cholesky (any diagonal pivoting, C or IC)
- L(i,j) (A(i,j) - ?kltj L(i,k)LT(k,j)) / L(j,j),
L(j,j) (A(j,j) - ?kltj L(j,k)LT(k,j))1/2 - L (and LT) plays triple role as a, b and c
40Some Corollaries of this lower bound (2/2)
words_moved ? (flops / M1/2 )
?
- Applies to simplified operations
- Ex Compute determinant of A(i,j) func(i,j)
using LU, where each func(i,j) called just
once so no inputs and 1 output - Still get ? (flops / M1/2 2n2) words_moved,
by imposing loads/stores
- Applies to compositions of operations
- Ex Compute Ak by repeated squaring, only input
A, output Ak - Still get ? (flops / M1/2 n2 log k ) by
imposing loads/stores, - Holds for any interleaving of operations
- Applies to some graph algorithms
- Ex All-Pairs-Shortest-Path by Floyd-Warshall
gijk and fij min - Get ? (F / M1/2) words_moved where F? n4 , n3
log n , n3
41Lower bounds for Orthogonal Transformations (1/4)
- Needed for QR, eig, SVD,
- Analyze Blocked Householder Transformations
- ?j1 to b (I 2 uj uTj ) I U T UT where U
u1 , , ub - Treat Givens as 2x2 Householder
- Model details and assumptions
- Write (I U T UT )A A U(TUT A) A UZ
where Z T(UT A) - Only count operations in all A UZ operations
- Generically a large fraction of the work
- Assume Forward Progress, that each successive
Householder transformation leaves previously
created zeros zero Ok for - QR decomposition
- Reductions to condensed forms (Hessenberg,
tri/bidiagonal) - Possible exception bulge chasing in banded case
- One sweep of (block) Hessenberg QR iteration
42Lower bounds for Orthogonal Transformations (2/4)
- Perform many A UZ where Z T(UT A)
- First challenge to applying theorem need to
collect all A-UZ into one big set to which model
(P) applies - Write them all as A(i,j) A(i,j) - ?k U(i,k)
Z(k,j) where k index of
k-th transformation, - k not necessarily index of column of A it comes
from - Second challenge All Z(k,j) may be S2/D2
- Recall S2/D2 means computed on the fly and
discarded - Ex An x 2n Qn x n Rn x 2n where 2n2 M so A
fits in cache - Represent Q as n(n-1)/2 2x2 Householder (Givens)
transforms - There are n2(n-1)/2 ?(M3/2) nonzero Z(k,j), not
O(M) - Still only do ?(M3/2) flops during segment
- But cant use Loomis-Whitney to prove it!
- Need a new idea
43Lower bounds for Orthogonal Transformations (3/4)
- Dealing with Z(k,j) being S2/D2
- How to bound flops in A(i,j) A(i,j) - ?k
U(i,k) Z(k,j) ? - Neither A nor U is S2/D2
- A either turned into R or U, which are output
- So at most 2M of each during segment
- flops ( U(i,k) ) ( columns A(,j) present
) - ( U(i,k) ) ( A(i,j) / min
nonzeros per column of A) - h O(M) / r
where h O(M) - How small can r be?
- To store h U(i,k) Householder vector entries
in r rows, there can be at most r in the
first column, r-1 in the second, etc., - (to maintain Forward Progress) so r(r-1)/2 ?
h so r ? h1/2 - flops h O(M) / r O(M) h1/2 O(M3/2) as
desired
44Lower bounds for Orthogonal Transformations (4/4)
- Theorem words_moved by QR decomposition using
(blocked) Householder transformations ?( flops
/ M1/2 ) - Theorem words_moved by reduction to Hessenberg
form, tridiagonal form, bidiagonal form, or one
sweep of QR iteration (or block versions of any
of these) ?( flops / M1/2 ) - Assuming Forward Progress (created zeros remain
zero) - Model Merge left and right orthogonal
transformations - A(i,j) A(i,j) - ?kLUL(i,kL) ZL(kL,j) -
?kRZR(i,kR) UR(j,kR)
45Minimizing latency requires new data structures
- To minimize latency, need to load/store whole
rectangular subblock of matrix with one message - Incompatible with conventional columnwise
(rowwise) storage - Ex Rows (columns) not in contiguous memory
locations - Blocked storage store as matrix of bxb blocks,
each block stored contiguously - Ok for one level of memory hierarchy, what if
more? - Recursive blocked storage store each block
using subblocks
46Blocked vs Cache-Oblivious Algorithms
- Blocked Matmul C AB explicitly refers to
subblocks of A, B and C of dimensions that depend
on cache size
Break Anxn, Bnxn, Cnxn into bxb blocks labeled
A(i,j), etc b chosen so 3 bxb blocks fit in
cache for i 1 to n/b, for j1 to n/b, for
k1 to n/b C(i,j) C(i,j) A(i,k)B(k,j)
b x b matmul another level of
memory would need 3 more loops
- Cache-oblivious Matmul C AB is independent of
cache
Function C MM(A,B) If A and B are 1x1 C
A B else Break Anxn, Bnxn, Cnxn into
(n/2)x(n/2) blocks labeled A(i,j), etc for
i 1 to 2, for j 1 to 2, for k 1 to
2 C(i,j) C(i,j) MM( A(i,k), B(k,j)
) n/2 x n/2 matmul
47Summary of dense sequential Cholesky algorithms
attaining communication lower bounds
- Algorithms shown minimizing Messages assume
(recursive) block layout - Many references (see reports), only some shown,
plus ours - Oldest reference may or may not include
analysis - Cache-oblivious are underlined, Green are ours,
? is unknown/future work - b block size, M fast memory size, n
dimension
Algorithm 2 Levels of Memory 2 Levels of Memory Multiple Levels of Memory Multiple Levels of Memory
Words Moved and Messages Words Moved and Messages
Cholesky LAPACK (with b M1/2) Gustavson, 97 recursive, assumed good MatMul available BDHS09 BDHS09 like Gustavson,97 Ahmed,Pingali,00 Uses recursive Cholesky, TRSM and Matmul (?same) BDHS09 has analysis (?same) BDHS09 has analysis
48Summary of dense sequential QR algorithms
attaining communication lower bounds
- Algorithms shown minimizing Messages assume
(recursive) block layout - Many references (see reports), only some shown,
plus ours - Oldest reference may or may not include
analysis - Cache-oblivious are underlined, Green are ours,
? is unknown/future work - b block size, M fast memory size, n
dimension
Algorithm 2 Levels of Memory 2 Levels of Memory Multiple Levels of Memory Multiple Levels of Memory
Words Moved and Messages Words Moved and Messages
QR Elmroth, Gustavson,98 Recursive, may do more flops LAPACK (only for n?M and bM1/2, or n2?M) DGHL08 Frens,Wise,03 explicit Q only, 3x more flops DGHL08 implicit Q, but represented differently, same flop count Elmroth, Gustavson,98 DGHL08? Frens,Wise,03 DGHL08? Can we get the same flop count?
49Summary of dense sequential LU algorithms
attaining communication lower bounds
- Algorithms shown minimizing Messages assume
(recursive) block layout - Many references (see reports), only some shown,
plus ours - Oldest reference may or may not include
analysis - Cache-oblivious are underlined, Green are ours,
? is unknown/future work - b block size, M fast memory size, n
dimension
Algorithm 2 Levels of Memory 2 Levels of Memory Multiple Levels of Memory Multiple Levels of Memory
Words Moved and Messages Words Moved and Messages
LU with pivoting Toledo, 97 recursive, assumes good TRSM and Matmul available LAPACK (only for n?M and bM1/2, or n2?M) GDX08 Different than partial pivoting, still stable GDX08 Toledo, 97 GDX08? GDX08?
50Same idea for LU, almost GDX08
- First idea Use same reduction tree as in QR,
with LU at each node - Numerically unstable!
- Need a different pivoting scheme
- At each node in tree, TSLU selects b pivot rows
from 2b candidates from its 2 child nodes - At each node, do LU on 2b original rows selected
by child nodes, not U factors from child
nodes - When TSLU done, permute b selected rows to top of
original matrix, redo b steps of LU without
pivoting - Thm Same results as partial pivoting on
different input matrix whose entries are the same
magnitudes as original - CALU Communication Avoiding LU for general A
- Use TSLU for panel factorizations
- Apply to rest of matrix
- Cost redundant panel factorizations
- Benefit
- Stable in practice, but not same pivot choice as
GEPP - b times fewer messages overall - faster
51Performance vs ScaLAPACK
- TSLU (Tall-Skinny)
- IBM Power 5
- Up to 4.37x faster (16 procs, 1M x 150)
- Cray XT4
- Up to 5.52x faster (8 procs, 1M x 150)
- CALU (Square)
- IBM Power 5
- Up to 2.29x faster (64 procs, 1000 x 1000)
- Cray XT4
- Up to 1.81x faster (64 procs, 1000 x 1000)
- See INRIA Tech Report 6523 (2008), paper at SC08
52CALU speedup prediction for a Petascale machine -
up to 81x faster
P 8192
Petascale machine with 8192 procs, each at 500
GFlops/s, a bandwidth of 4 GB/s.
53Eigenvalue problems and SVD
- Uses spectral divide and conquer
- Idea goes back to Bulgakov, Godunov, Malyshev
- Roughly If A2k QR, leading columns of Q
converge to span invariant subspace for
eigenvalues of A outside unit circle so QAQ
becomes block upper triangular - Repeated squaring done implicitly
- Bai, D., Gu, 97
- Reduced just to Matmul, QR, and Generalized QR
with pivoting - Eliminated need for exp(A), other improvements
- DDH,07
- Introduced randomized rank revealing QR
- Q,Rqr(randn(n,n)), Q,Rqr(AQ)
- Just need Matmul and QR
- BDD,09
- Shows that randomized GQR also rank revealing
- Shows limits of divide-and-conquer, depending on
pseudospectrum - Cache-oblivious, using Frens,Wise,03
54Implicit Repeated Squaring
for each j,
55Sparse Linear Algebra
- Thm (George 73, Hoffman/Martin/Rose 73,
Eisenstat/ Schultz/Sherman 76 ) Cholesky on an
ns x ns sparse matrix whose graph is an
s-dimensional nearest neighbor grid does ?(
n3(s-1) ) flops attainable by nested dissection
ordering - Proof most work in dense Cholesky on final
submatrix of size ?(ns-1) - Corollary BDHS09,G09 Sequential Sparse
Cholesky on such a matrix moves ?( n3(s-1) /
M1/2 ) words, and Parallel Sparse Cholesky moves
?( n2(s-1) / (P log n )1/2 ) words on P procs
- Assumes 2D parallel algorithm, i.e. minimal
memory - Thm G09. Attainable in parallel case by nested
dissection, eg with PSPASES.
56Locally Dependent Entries for x,Ax, A
tridiagonal, 2 processors
Proc 1
Proc 2
Can be computed without communication
57Locally Dependent Entries for x,Ax,A2x, A
tridiagonal, 2 processors
Proc 1
Proc 2
Can be computed without communication
58Locally Dependent Entries for x,Ax,,A3x, A
tridiagonal, 2 processors
Proc 1
Proc 2
Can be computed without communication
59Locally Dependent Entries for x,Ax,,A4x, A
tridiagonal, 2 processors
Proc 1
Proc 2
Can be computed without communication
60Locally Dependent Entries for x,Ax,,A8x, A
tridiagonal, 2 processors
Proc 1
Proc 2
Can be computed without communication k8 fold
reuse of A
61Remotely Dependent Entries for x,Ax,,A8x, A
tridiagonal, 2 processors
Proc 1
Proc 2
One message to get data needed to compute
remotely dependent entries, not k8 Minimizes
number of messages latency cost Price
redundant work ? surface/volume ratio
62Remotely Dependent Entries for x,Ax,A2x,A3x, A
irregular, multiple processors
63Sequential x,Ax,,A4x, with memory hierarchy
v
One read of matrix from slow memory, not
k4 Minimizes words moved bandwidth cost No
redundant work
64Performance results on 8-Core Clovertown
65Direct linear algebra Summary of lower bounds
- Sequential case with fast memory of size M
- Lower bound on words moved to/from slow memory
? ( flops / M1/2 ) - Ex Dense O(n3) MatMul ? (n3 / M1/2 )
- Lower bound on messages to/from slow memory
? ( flops / M3/2 ) - Ex Dense O(n3) MatMul ? (n3 / M3/2 )
- Parallel case on P processors
- Let NNZ be total memory needed assume balanced
- Lower bound on words communicated
? ( flops / (P NNZ )1/2 ) - Ex Dense O(n3) MatMul (so NNZ 3n2) ? (n2
/ P1/2 ) - Lower bound on messages sent
? ( flops P1/2 / NNZ3/2 ) - Ex Dense O(n3) MatMul ? (P1/2 )
66Naïve Matrix Multiply
- implements C C AB
- for i 1 to n
- for j 1 to n
- for k 1 to n
- C(i,j) C(i,j) A(i,k) B(k,j)
A(i,)
C(i,j)
C(i,j)
B(,j)
67Naïve Matrix Multiply
- implements C C AB
- for i 1 to n
- read row i of A into fast memory
- for j 1 to n
- read C(i,j) into fast memory
- read column j of B into fast memory
- for k 1 to n
- C(i,j) C(i,j) A(i,k) B(k,j)
- write C(i,j) back to slow memory
Number of slow memory references m n3 to
read each column of B n times n2
to read each row of A once 2n2 to
read and write each element of C once
n3 3n2 Similar for any other order of 3 loops
68Blocked (Tiled) Matrix Multiply
- Consider A,B,C to be N-by-N matrices of b-by-b
blocks where bn / N is called
the block size - for i 1 to N
- for j 1 to N
- read block C(i,j) into fast memory
- for k 1 to N
- read block A(i,k) into fast
memory - read block B(k,j) into fast
memory - C(i,j) C(i,j) A(i,k)
B(k,j) do a matrix multiply on blocks - write block C(i,j) back to slow memory
A(i,k)
C(i,j)
C(i,j)
B(k,j)
69Blocked (Tiled) Matrix Multiply
- Recall
- m is amount memory traffic between slow and
fast memory - matrix has nxn elements, and NxN blocks each
of size bxb
- m Nn2 read each block of B N3 times
(N3 b2 N3 (n/N)2 Nn2) - Nn2 read each block of A N3
times - 2n2 read and write each block
of C once - (2N 2) n2
- ? 2 n3 / b smaller than naïve by a
factor of b - The bigger the b the better How big can we make
b? - Limit need to fit 3 b x b submatrices into fast
memory of size M, so 3b2 ? M - So cant reduce m below ? (n3 / M1/2) by this
approach - Is there any way to make m smaller?
70Limits to Optimizing Matrix Multiply
- The blocked algorithm changes the order in which
values are accumulated into each Ci,j by
applying
Commutativity and Associativity of Addition (CAA) - Get slightly different answers from naïve code,
because of roundoff - OK - Theorem (Hong Kung, 1981) Any reorganization
of this algorithm (that uses only CAA) on a
machine with fast memory size M does at least
memory references ? (n3 / M1/2) - Attained by blocked matmul
- Theorem (Irony, Toledo Tiskin, 2004) Suppose
we are doing parallel matrix multiplication on P
processors, where each processor can store O(n2 /
P) words. Then for any algorithm (using only
CAA), at least one processor needs to communicate
at least O(n2 / P1/2) words - Attained by Cannons algorithm
- Extends to 3D matmul using P1/3 times as much
memory as minimal
71Some Corollaries (1/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 1 Dense or sparse matrix-multiply C
AB requires at least multiplies/ (8M)1/2 M
slow memory references
- Not always attainable (suppose A, B diagonal)
- Tighter lower bound max of above and total
input/output size
72Some Corollaries (2/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 2 Dense or sparse matrix-multiply C
AB - on P parallel processors,
- where each processor stores
- NNZ/P ( nnz(A) nnz(B) nnz(C) ) / P words
- requires at least (multiplies/P)/(8
NNZ/P)1/2 NNZ/P (multiplies)/(8 NNZ P)1/2
NNZ/P words communicated
73Some Corollaries (3/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 3 Corollaries 1 and 2 apply to ATA and
A2. - Theorem permits a(i,k) and b(k,j) to overlap
74Some Corollaries (4/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 4 Corollaries 1 and 2 apply to TRSM C
A-1B when A triangular (dense or sparse,
sequential or parallel) - C(i,j) (B(i,j) - ?klti A(i,k)C(k,j)) / A(i,i)
A lower tri. - Matches (P)
- C plays double role as b and c
- Subtraction from B(i,j), division by A(i,i)
some other arguments - Lower bound on all reorderings, even invalid ones
75Attaining Bandwidth Lower Bound for Dense TRSM
- function X R_TRSM(A,B) . Compute X A-1B
- if dimension 1
- return X B/A
- else
- partition B B11 , B12 B21 , B22 etc
- X11 R_TRSM( A11 , B11 )
- X21 R_TRSM( A22 , B21 A21 X11 )
- need to do matmul right
- X12 R_TRSM( A11 , B12 )
- X22 R_TRSM( A22 , B22 A21 X12 )
- need to do matmul right
- Cache-oblivious approach
- Leiserson, Frigo, Prokop, Blelloch, Pingali,
- Minimizes bandwith across multiple levels of
memory hierarchy - What about latency?
76Some Corollaries (5/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 5 Corollaries 1 and 2 apply to LU
factorization (dense or sparse, any pivoting, LU
or ILU, seq. or par.) - L(i,j) (A(i,j) - ?kltj L(i,k)U(k,j)) / U(j,j)
- U(i,j) A(i,j) - ?klti L(i,k)U(k,j)
- Matches (P)
- L (and U) plays double role as c and a (c and b)
- Subtractions, division by U(j,j) some other
arguments - Lower bound on all reorderings, even invalid ones
77Some Corollaries (6/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 6 Corollaries 1 and 2 apply to
Cholesky (dense or sparse, any diag.
pivoting, C or IC, seq. or par.) - L(i,j) (A(i,j) - ?kltj L(i,k)LT(k,j)) / L(j,j)
- L(j,j) (A(j,j) - ?kltj L(j,k)LT(k,j))1/2
- Matches (P)
- L (and LT) plays triple role as a, b and c
- Subtractions, division, square root some other
arguments - Lower bound on all reorderings, even invalid ones
78Some Corollaries (7/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Corollary 7 Sparse Cholesky, where graph of
matrix contains an n x n mesh, requires ?(n3
/M1/2 ) slow mem refs - Eg 5 point Laplacian on n x n mesh
- Combine Corollary 6 with lower bound on flops by
George , Hoffman/Martin/Rose
79Some Corollaries (8/8)
- Theorem To evaluate (P) with fast memory of size
M requires at least gijk s / (8M)1/2 M
slow memory refs
- Conjecture 8 Corollaries 1 and 2 apply to QR
(dense or sparse, any pivoting, seq.
or par.) - Proof of dense case only in EECS Tech Report
2008-89
80Do any LU/QR algs reach lower bounds?
- Only dense case understood
- LAPACK attains neither O((n2/M)1/2) x more
bandwidth - Recursive seq. algs minimize bandwidth, not
latency - Toledo for LU, Elmroth/Gustavson for QR
- ScaLAPACK attains bandwidth lower bound
- But sends O((n2/P)1/2) x more messages
- New LU and QR algorithms do attain both lower
bounds, both seq. and par. - LU need to abandon partial pivoting (but still
stable) - QR need to represent Q as tree of Householder
matrices - Neither new alg works for multiple mem hierarchy
levels - See EECS TR 2008-89 for lots of related work
- Oldest Golub/Plemmons/Sameh (1988)
81Minimizing Comm. in Parallel QR
- QR decomposition of m x n matrix W, m gtgt n
- TSQR Tall Skinny QR
- P processors, block row layout
- Usual Parallel Algorithm
- Compute Householder vector for each column
- Number of messages ? n log P
- Communication Avoiding Algorithm
- Reduction operation, with QR as operator
- Number of messages ? log P
82TSLU LU factorization of a tall skinny matrix
First try the obvious generalization of TSQR
83Growth factor for TSLU based factorization
- Unstable for large P and large matrices.
- When P rows, TSLU is equivalent to parallel
pivoting.
Courtesy of H. Xiang
84Growth factor for better CALU approach
Like threshold pivoting with worst case threshold
.33 , so L lt 3 Testing shows about same
residual as GEPP
85Do any Cholesky algs reach lower bounds?
- Only dense case understood
- LAPACK (with right block size) or recursive
Cholesky minimize bandwidth - Recursive Ahmed/Pingali, Gustavson/Jonsson,
Andersen/ Gustavson/Wasniewski, Simecek/Tvrdik, a
la Toledo - LAPACK can minimize latency with blocked data
structure - Ahmed/Pingali minimize bandwidth and latency
across multiple levels of memory hierarchy - Simultaneously minimize communication between all
pairs L1/L2/L3/DRAM/disk/ - Cache-oblivious, space-filling curve layout
- ScaLAPACK minimizes bandwidth and latency (mod
polylog P) - Need right choice of block size
- Details in EECS TR 2009-29
86Future work (1/2)
- Extend lower bound to QR (soon!)
- Dense sequential QR and LU that minimize
bandwidth and latency over multiple memory
hierarchy levels - Extend to fast algorithms (see next slide)
- Extend to eigenvalue problems (see next slide)
- Extend to 3D algorithms
- Attaining upper bounds for sparse matrices?
- Need to examine special cases of interest
- Extend to heterogeneous architectures
- Layers of parallelism and memory hiearchy
- Varying flops speeds, bandwidths, latencies
- Autotuning
- Redo all of LAPACK and ScaLAPACK
87Future Work (2/2) Asymptotically Faster
Algorithms
- Thm (D., Dumitriu, Holtz, Kleinberg) (Numer.Math.
2007) - Given any matmul running in O(n?) ops for some
?gt2, it can be made stable and still run in
O(n??) ops, for any ?gt0. - Strassen already stable
- Current record ? ? 2.38
- Thm (D., Dumitriu, Holtz) (Numer. Math. 2008)
- Given any stable matmul running in O(n??) ops,
it is possible to do backward stable dense linear
algebra in O(n??) ops - GEPP, QR
- rank revealing QR (randomized)
- (Generalized) Schur decomposition, SVD
(randomized) - Also reduces communication to O(n??)
- But lower bounds?
- But constants?