I tried running to run a toy example for some matrix multiplication and addition using distributed TensorFlow.
My goal was to compute (A^n + B^n)
where A[,]
and B[,]
are LxL
matrices.
I used 2 machines on the public cloud to compute A^n
on one machine and B^n
on the second machine, than the addition again on the first machine.
When the machines had CPU only - my script worked great.
When both had GPU - It failed to run in a reasonable time! It had a huge latency...
My question - what did I do wrong in my script?
Note that for machine2 ( task:1
) I used server.join()
and I used machine 1 ( task:0
) as the client in this with-in graph.
#------------------------------------------------------------------
from zmq import Stopwatch; aClk_E2E = Stopwatch(); aClk_E2E.start()
#------------------------------------------------------------------
from __future__ import print_function
import numpy as np
import tensorflow as tf
import datetime
IP_1 = '10.132.0.2'; port_1 = '2222'
IP_2 = '10.132.0.3'; port_2 = '2222'
cluster = tf.train.ClusterSpec( { "local": [ IP_1 + ":" + port_1,
IP_2 + ":" + port_2
],
}
)
server = tf.train.Server( cluster,
job_name = "local",
task_index = 0
)
# server.join() # @machine2 ( task:1 )
n = 5
L = 1000
def matpow( M, n ):
if n < 1: # Abstract cases where n < 1
return M
else:
return tf.matmul( M, matpow( M, n - 1 ) )
G = tf.Graph()
with G.as_default():
with tf.device( "/job:local/task:1/cpu:0" ):
c1 = []
tB = tf.placeholder( tf.float32, [L, L] ) # tensor B placeholder
with tf.device( "/job:local/task:1/gpu:0" ):
c1.append( matpow( tB, n ) )
with tf.device( "/job:local/task:0/cpu:0" ):
c2 = []
tA = tf.placeholder( tf.float32, [L, L] ) # tensor A placeholder
with tf.device( "/job:local/task:0/gpu:0" ):
c2.append( matpow( tA, n ) )
sum2 = tf.add_n( c1 + c2 )
#---------------------------------------------------------<SECTION-UNDER-TEST>
t1_2 = datetime.datetime.now()
with tf.Session( "grpc://" + IP_1 + ":" + port_1, graph = G ) as sess:
A = np.random.rand( L, L ).astype( 'float32' )
B = np.random.rand( L, L ).astype( 'float32' )
sess.run( sum2, { tA: A, tB: B, } )
t2_2 = datetime.datetime.now()
#---------------------------------------------------------<SECTION-UNDER-TEST>
#------------------------------------------------------------------
_ = aClk_E2E.stop()
#------------------------------------------------------------------
print( "Distributed Computation time: " + str(t2_2 - t1_2))
print( "Distributed Experiment took: {0: > 16d} [us] End-2-End.".format( _ ) )
A first step into this domain is always challenging. Lost certainties, that were taken as granted in monolythic computing previous experience, many new challenges, that had no similar issues in single-node process coordinations, many new surprises from new orders of magnitude of distributed execution timing- and distributed ( coordination, if not dead-lock and/or live-lock ) blocking-issues.
Thanks for added some quantitative facts ~ 15 seconds being "too-long" for A[1000,1000];B[1000,1000];n=5
-- so far so good.
Would you mind to add the above proposed code-changes and re-run the experiment on the same real-infrastructure?
This will help the rest of this started work ( WIP here ).
-- THANKS IN ADVANCE to run + post updated facts.
It is hard to continue ATM with quantitatively supported statements, however, my gut feeling suspect is in this one:
def matpow( M, n ):
return M if ( n < 1 ) else tf.matmul( M, matpow( M, n - 1 ) )
It uses a recursion, that may go too deep for GPU-cross-compiler/assembler analysers and given scales of tensors, the GPU SMX that are "fast" for mathematically dense-kernel microcode, the SMX-local SM_registers ( having latencies of about a 22 GPU_CLK
s ( well, may just 8, if smart-optimised to get fetched from well LRU-aligned SM_L1-cache-line ) will have to spill-over to global_MEMORY ( as each SM has a chance to store way less than a 1 KB in repetitive memory-access latency-most-friendly SM_Registers, but matmul() does never re-use any cell of a matrix so latency-hiding will never pay less than each global_MEMORY access + the [PSPACE]
scales .. ), where suddenly experience about a 600+ GPU_CLK
s latency penalties.
While this narrated animation of HPC matmul()
mentions typical CPU / Lx-cache / Memory hierarchy, the message why any O(N^3)
processing must get incredibly slow on GPU-s for N
going larger than could fit onto SM_registers is yet easily visible ( and as you see it, imagine all your cache-friendliness lost in recursion matpow()
).
GPU-kernels get best results for small-scale still-SM-local-convolutions ( where right this SMX-locality allows for good [Data:SMX-local SM_REG] alignment ( and zero cross-SMX-communication need to take place ( which is not the case in matmul() processing anywhere larger than just ~ 7 x 7 matrices, that could fit on SM_REG-silicon and anything above this has to start super-smart stencil-aligned gymnastics, if striving pay the minimum necessary sum of just-GPU-local memIO-latencies ( an the story goes forwards, if poor host2dev / dev2host IO-takes place, and many more places, where the execution-performance is uncontrollably poor ) ).
The latency costs even for a single matmul( A, A )
are suddenly hell-loooong, of compared to typical image-specialised kernels. ( Sure, there are advanced techniques, how to by-pass this specialised-silicon limits, but even if matmul()
would have been a master of top tier HPC-block-matrix-ops, it will be lest as a naive-matmul()
once recursive calls get on the stage -- that will kill even the smarted tricks, as there is no [SPACE]
for the "stack"-intermediate-values, for which the auto-generated-kernel-code will pay in immense [TIME]
penalties .. even for such small scales, like 1000x1000 ).
Category GPU
| Hardware
| Unit
| | Throughput
| | | Execution
| | | Latency
| | | | PTX instructions Note
|____________________________|____________|_______________|__________________|_____________________________________________________________________|________________________________________________________________________________________________________________________
Load_shared LSU 2 + 30 ld, ldu Note, .ss = .shared ; .vec and .type determine the size of load. Note also that we omit .cop since no cacheable in Ocelot
Load_global LSU 2 + 600 ld, ldu, prefetch, prefetchu Note, .ss = .global; .vec and .type determine the size of load. Note, Ocelot may not generate prefetch since no caches
Load_local LSU 2 + 600 ld, ldu, prefetch, prefetchu Note, .ss = .local; .vec and .type determine the size of load. Note, Ocelot may not generate prefetch since no caches
Load_const LSU 2 + 600 ld, ldu Note, .ss = .const; .vec and .type determine the size of load
Load_param LSU 2 + 30 ld, ldu Note, .ss = .param; .vec and .type determine the size of load
| |
Store_shared LSU 2 + 30 st Note, .ss = .shared; .vec and .type determine the size of store
Store_global LSU 2 + 600 st Note, .ss = .global; .vec and .type determine the size of store
Store_local LSU 2 + 600 st Note, .ss = .local; .vec and .type determine the size of store
Read_modify_write_shared LSU 2 + 600 atom, red Note, .space = shared; .type determine the size
Read_modify_write_global LSU 2 + 600 atom, red Note, .space = global; .type determine the size
| |
Texture LSU 2 + 600 tex, txq, suld, sust, sured, suq
| |
Integer ALU 2 + 24 add, sub, add.cc, addc, sub.cc, subc, mul, mad, mul24, mad24, sad, div, rem, abs, neg, min, max, popc, clz, bfind, brev, bfe, bfi, prmt, mov
| | Note, these integer inst. with type = { .u16, .u32, .u64, .s16, .s32, .s64 };
| |
Float_single ALU 2 + 24 testp, copysign, add, sub, mul, fma, mad, div, abs, neg, min, max Note, these Float-single inst. with type = { .f32 };
Float_double ALU 1 + 48 testp, copysign, add, sub, mul, fma, mad, div, abs, neg, min, max Note, these Float-double inst. with type = { .f64 };
Special_single SFU 8 + 48 rcp, sqrt, rsqrt, sin, cos, lg2, ex2 Note, these special-single with type = { .f32 };
Special_double SFU 8 + 72 rcp, sqrt, rsqrt, sin, cos, lg2, ex2 Note, these special-double with type = { .f64 };
|
Logical ALU 2 + 24 and, or, xor, not, cnot, shl, shr
Control ALU 2 + 24 bra, call, ret, exit
|
Synchronization ALU 2 + 24 bar, member, vote
Compare & Select ALU 2 + 24 set, setp, selp, slct
|
Conversion ALU 2 + 24 Isspacep, cvta, cvt
Miscellanies ALU 2 + 24 brkpt, pmevent, trap
Video ALU 2 + 24 vadd, vsub, vabsdiff, vmin, vmax, vshl, vshr, vmad, vset
The technical post webpages of this site follow the CC BY-SA 4.0 protocol. If you need to reprint, please indicate the site URL or the original address.Any question please contact:yoyou2525@163.com.