Download CUDA Restrictions – Task Parallelism

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Wrapup and Open Issues
Kevin Skadron
University of Virginia Dept. of Computer Science
LAVA Lab
Outline
Objective of this segment: explore interesting open
questions, research challenges
CUDA restrictions (we have covered most of these)
CUDA ecosystem
CUDA as a generic platform for manycore research
Manycore architecture/programming in general
© Kevin Skadron, 2008
CUDA Restrictions – Thread Launch
Threads may only be created by launching a kernel
Thread blocks run to completion—no provision to
yield
Thread blocks can run in arbitrary order
No writable on-chip persistent state between thread
blocks
Together, these restrictions enable
Lightweight thread creation
Scalable programming model (not specific to number of
cores)
Simple synchronization
© Kevin Skadron, 2008
CUDA Restrictions – Concurrency
Task parallelism is restricted
Global communication is restricted
No writable on-chip persistent state between thread
blocks
Recursion not allowed
Data-driven thread creation, and irregular fork-join
parallelism are therefore difficult
Together, these restrictions allow focus on
perf/mm2
scalability
© Kevin Skadron, 2008
CUDA Restrictions – GPU-centric
No OS services within a kernel
e.g., no malloc
CUDA is a manycore but single-chip solution
Multi-chip, e.g. multi-GPU, requires explicit management
in host program, e.g. separately managing multiple CUDA
devices
Compute and 3D-rendering can’t run concurrently
(just like dependent kernels)
© Kevin Skadron, 2008
CUDA Ecosystem Issues
These are really general parallel-programming ecosystem
issues
#1 issue: need parallel libraries and skeletons
Ideally the API is platform independent
#2 issue: need higher-level languages
Simplify programming
Provide information about desired outcomes, data structures
May need to be domain specific
Allow underlying implementation to manage parallelism, data
Examples: D3D, MATLAB, R, etc.
#3 issue: debugging for correctness and performance
CUDA debugger and profiler in beta, but…
Even if you have the mechanics, it is an information visualization
challenge
GPUs do not have precise exceptions
© Kevin Skadron, 2008
Outline
CUDA restrictions
CUDA ecosystem
CUDA as a generic platform for manycore research
Manycore architecture/programming in general
© Kevin Skadron, 2008
CUDA for Architecture Research
CUDA is a promising vehicle for exploring many aspects of
parallelism at an interesting scale on real hardware
CUDA is also a great vehicle for developing good parallel
benchmarks
What about for architecture research?
We can’t change the HW
CUDA is good for exploring bottlenecks
Programming model: what is hard to express, how could
architecture help?
Performance bottlenecks: where are the inefficiencies in real
programs?
But how to evaluate benefits of fixing these bottlenecks?
Open question
Measure cost at bottlenecks, estimate benefit of a solution?
– Will our research community accept this?
© Kevin Skadron, 2008
Programming Model
General manycore challenges…
Balance flexibility and abstraction vs. efficiency and scalability
MIMD vs SIMD, SIMT vs. vector
Barriers vs. fine-grained synchronization
More flexible task parallelism
High thread count requirement of GPUs
Balance ease of first program against efficiency
Software-managed local store vs. cache
Coherence
Genericity vs. ability to exploit fixed-function hardware
Ex: OpenGL exposes more GPU hardware that can be used to
good effect for certain algorithms—but harder to program
Balance ability to drill down against portability, scalability
Control thread placement
Challenges due to high off-chip offload costs
Fine-grained global RW sharing
Seamless scalability across multiple chips (“manychip”?),
distributed systems
© Kevin Skadron, 2008
Scaling
How do we use all the transistors?
ILP
More L1 storage?
More L2?
More PEs per “core”?
More cores?
More true MIMD?
More specialized hardware?
How do we scale when we run into the power wall?
© Kevin Skadron, 2008
Thank you
Questions?
© Kevin Skadron, 2008
CUDA Restrictions – Task Parallelism
(extended)
Task parallelism is restricted and can be awkward
Hard to get efficient fine-grained task parallelism
Between threads: allowed (SIMT), but expensive due to
divergence and block-wide barrier synchronization
Between warps: allowed, but awkward due to block-wide barrier
synchronization
Between thread blocks: much more efficient, code still a bit
awkward (SPMD style)
Communication among thread blocks inadvisable
Communication among blocks generally requires new kernel call,
i.e. global barrier
Coordination (i.e. shared queue pointer) ok
No writable on-chip persistent state between thread blocks
These restrictions stem from focus on
perf/mm2
support scalable number of cores
© Kevin Skadron, 2008
CUDA Restrictions – HW Exposure
CUDA is still low-level, like C – good and bad
Requires/allows manual optimization
Manage concurrency, communication, data transfers
Manage scratchpad
Performance is more sensitive to HW characteristics than C on a
uniprocessor (but this is probably true of any manycore system)
Thread count must be high to get efficiency on GPU
Sensitivity to number of registers allocated
– Fixed total register file size, variable registers/thread, e.g. G80: #
registers/thread limits # threads/SM
– Fixed register file/thread, e.g., Niagara:
small register allocation wastes register file space
– SW multithreading: context switch cost f(# registers/thread)
Memory coalescing/locality
…These issues will arise in most manycore architectures
Many more interacting hardware sensitivities (# regs/thread, #
threads/block, shared memory usage, etc.)—performance tuning
is tricky
Need libraries and higher-level APIs that have been optimized
to account for all these factors
© Kevin Skadron, 2008
SIMD Organizations
SIMT: independent threads, SIMD hardware
First used in Solomon and Illiac IV
Each thread has its own PC and stack, allowed to follow
unique execution path, call/return independently
Implementation is optimized for lockstep execution of these
threads (32-wide warp in Tesla)
Divergent threads require masking, handled in HW
Each thread may access unrelated locations in memory
Implementation is optimized for spatial locality—adjacent
words will be coalesced into one memory transaction
Uncoalesced loads require separate memory transactions,
handled in HW
Datapath for each thread is scalar
© Kevin Skadron, 2008
SIMD Organizations (2)
Vector
First implemented in CDC, commercial big time in
MMX/SSE/Altivec/etc.
Multiple lanes handled with a single PC and stack
Divergence handled with predication
Promotes code with good lockstep properties
Datapath operates on vectors
Computing on data that is non-adjacent from memory
requires vector gather if available, else scalar load and
packing
Promotes code with good spatial locality
Chief difference: more burden on software
What is the right answer?
© Kevin Skadron, 2008
Backup
© Kevin Skadron, 2008
Other Manycore PLs
Data-parallel languages (HPF, Co-Array Fortran, various dataparallel C dialects)
DARPA HPCS languages (X10, Chapel, Fortress)
OpenMP
Titanium
Java/pthreads/etc.
MPI
Streaming
Key CUDA differences
Virtualizes PEs
Supports offload
Scales to large numbers of threads
Allows shared memory between (subsets of) threads
© Kevin Skadron, 2008
Persistent Threads
You can cheat on the rule that thread blocks should
be independent
Instantiate # thread blocks = # SMs
Now you don’t have to worry about execution order
among thread blocks
Lose hardware global barrier
Need to parameterize code so that it is not tied to a
specific chip/# SMs
© Kevin Skadron, 2008
DirectX – good case study
High-level abstractions
Serial ordering among primitives
= implicit synchronization
No guarantees about ordering within primitives
= no fine-grained synchronization
Domain-specific API is convenient for programmers and provides
lots of semantic information to middleware: parallelism, load
balancing, etc.
Domain-specific API is convenient for hardware designers: API has
evolved while underlying architectures have been radically different
from generation to generation and company to company
Similar arguments apply to Matlab, SQL, Map-Reduce, etc.

I’m not advocating any particular API, but these examples show that
high-level, domain-specific APIs are commercially viable and
effective in exposing parallelism
Middleware (hopefully common) translates domain-specific APIs to
general-purpose architecture that supports many different app. domains
© Kevin Skadron, 2008
HW Implications of Abstractions
If we are developing high-level abstractions and supporting
middleware, low-level macro-architecture is less important
Look at the dramatic changes in GPU architecture under DirectX
If middleware understands triangles/matrices/graphs/etc., it can
translate them to SIMD or anything else
HW design should focus on:
Reliability
Scalability (power, bandwidth, etc.)
Efficient support for important parallel primitives
– Scan, sort, shuffle, etc.
– Memory model
– SIMD: divergent code – work queues, conditional streams,
etc.
– Efficient producer-consumer communication
Flexibility (need economies of scale)
Need to support legacy code (MPI, OpenMP, pthreads) and
other “low-level languages” (X10, Chapel, etc.)
Low-level abstractions cannot cut out these users
Programmers must be able to “drill down” if necessary
© Kevin Skadron, 2008