CS 395 Last Lecture

Summary, Anti-summary, and Final Thoughts

2

Summary (1) Architecture

• Modern architecture designs are driven by energy constraints

• Shortening latencies is too costly, so we use parallelism in hardware to increase potential throughput

• Some parallelism is implicit (out-of-order superscalar processing,) but have limits

• Others are explicit (vectorization and multithreading,) and rely on software to unlock

3

Summary (2) Memory

• Memory technologies trade off energy and cost for capacity, with SRAM registers on one end and spinning platter hard disks on the other

• Locality (relationships between memory accesses) can help us get the best of all cases

• Caching is the hardware-only solution to capturing locality, but software-driven solutions exist too (memcache for files, etc.)

4

Summary (3) Software

• Want to fully occupy your hardware?– Express locality (tiling)– Vectorize (compiler or manual)– Multithread (e.g. OpenMP)– Accelerate (e.g. CUDA, OpenCL)

• Take the cost into consideration. Unless you’re optimizing in your free time, your time isn’t free.

5

Research Perspective (2010)

• Can we generalize and categorize the most important, generally applicable GPU Computing software optimizations?– Across multiple architectures– Across many applications

• What kinds of performance trends are we seeing from successive GPU generations?

• Conclusion – GPUs aren’t special, and parallel programming is getting easier

6

Application Survey

• Surveyed the GPU Computing Gems chapters• Studied the Parboil benchmarks in detail

Results: • Eight (for now) major categories of

optimization transformations– Performance impact of individual optimizations on

certain Parboil benchmarks included in the paper

1: (Input) Data Access Tiling

7

DRAM

DRAM

Cache

DRAM

Scratchpad

ExplicitCopy

ImplicitCopy

LocalAccess

LocalAccess

8

2. (Output) Privatization

• Avoid contention by aggregating updates locally

• Requires storage resources to keep copies of data structures

PrivateResults

LocalResults

GlobalResults

9

Running Example: SpMV

Ax = v

Row

Data

Col

vx

A

10

Running Example: SpMV

Ax = v

Row

Data

Col

A

vx

11

3. “Scatter to Gather” Transformation

Ax = v v

Row

Data

Col

A

x

12

3. “Scatter to Gather” Transformation

Ax = v v

Row

Data

Col

A

x

13

4. Binning

A

14

5. Regularization (Load Balancing)

15

6. Compaction

16

7. Data Layout Transformation

17

7. Data Layout Transformation

18

8. Granularity Coarsening• Parallel execution often requires redundant and

coordination work– Merging multiple threads into one allows reuse of result,

reducing redundancy

Essential

Redundant

4-wayparallel

2-wayparallel

Time

How much faster do applications really get each hardware generation?

20

Unoptimized Code Has Improved Drastically

• Orders of magnitude speedup in many cases

• Hardware does not solve all problems– Coalescing (lbm)– Highly contentious

atomics (bfs)

21

Optimized Code Is Improving Faster than “Peak Performance”

• Caches capture locality scratchpad can’t efficiently (spmv, stencil)

• Increased local storage capacity enables extra optimization (sad)

• Some benchmarks need atomic throughput more than flops (bfs, histo)

22

Optimization Still Matters• Hardware never

changes algorithmic complexity (cutcp)

• Caches do not solve layout problems for big data (lbm)

• Coarsening still makes a big difference (cutcp, sgemm)

• Many artificial performance cliffs are gone (sgemm, tpacf, mri-q)

23

Stuff we haven’t covered

• Good tools out there for profiling code beyond good timing (cache misses, etc.) If you can’t find why a particular piece of code is taking so long, look into hardware performance counters.

• Patterns and practice– Some of the major patterns of optimization we

covered, but only the basic ones. Many optimization patterns are algorithmic.

24

Fill Out Evaluations!