Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
146 views

Accelerating CUDA Graph Algorithms at Maximum Warp

This document summarizes a research paper about accelerating CUDA graph algorithms at maximum warp. [1] It presents a warp-centric programming method to address workload imbalance and memory inefficiency issues that arise when directly mapping graph algorithms to GPU architectures. [2] The method runs programs in either SISD or SIMT phases to better utilize hardware and avoid path divergence. [3] Experimental results show this approach outperforms traditional thread-to-data mappings, especially on irregular graphs.

Uploaded by

thangmle
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
146 views

Accelerating CUDA Graph Algorithms at Maximum Warp

This document summarizes a research paper about accelerating CUDA graph algorithms at maximum warp. [1] It presents a warp-centric programming method to address workload imbalance and memory inefficiency issues that arise when directly mapping graph algorithms to GPU architectures. [2] The method runs programs in either SISD or SIMT phases to better utilize hardware and avoid path divergence. [3] Experimental results show this approach outperforms traditional thread-to-data mappings, especially on irregular graphs.

Uploaded by

thangmle
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 25

Accelerating CUDA Graph Algorithms at Maximum Warp

By: Sungpack Hong, Sang Kyun Kim, Tayo Oguntebi, Kunle Olukotun

Presenter: Thang M. Le

Authors

Sungpack Hong

Ph.D graduate from Stanford Currently, a principal member of technical staff at Oracle

Sang Kyun Kim

Ph.D candidate at Stanford

Tayo Oguntebi

Ph.D candidate at Stanford

Kunle Olukotun Professor of Electrical Engineering & Computer Science at Stanford Director of Pervasive Parallel Laboratory

Agenda

What Is the Problem? Why Does the Problem Exist? Warp-Centric Programming Method Other Techniques Experimental Results Study of Architectural Effects Q&A

What Is the Problem?

The Parallel Random Access Machine (PRAM) abstract is often used to investigate theoretical parallel performance of graph algorithm PRAM approximation is quite accurate in supercomputer domains such as Cray XMT PRAM based algorithms fail to perform well on GPU architectures due to the workload imbalance among threads

Why Does the Problem Exist?

CUDA thread model exhibits certain discrepancies with the GPU architecture

Notably, no explicit notion of warps Different behaviors when accessing memory


Requests targeting the same address are merged Spatial locality accesses are maximally coalesced All other memory requests are serialized

Why Does the Problem Exist?

SIMT eliminates SIMD constraints by allowing threads in a warp to pursue on different paths (aka path divergence) Path divergence provides more flexibilities at the cost of performance Path divergence leads to hardware underutilization

Scattering memory access patterns

Why Does the Problem Exist?

A thread that processes a high-degree node will iterate the loop at line 23 many more times than other threads, stalling other threads in the warp
Path divergence

Why Does the Problem Exist?

Non-coalesced memory

Warp-Centric Programming Method

A program can be run in either SISD phase or SIMT phase:

SISD:

all threads in a warp are executed on the same data degree of parallelism (per SM) = O(# concurrent warps) Each thread is executed on different data degree of parallelism (per SM) = O(# threads in a warp x # concurrent warps)

SIMT:

By default, all threads in a warp are executed following SISD When appropriate, switch to SIMT to exploit data parallelism

Warp-Centric Programming Method

Is it safe to run a logic in SISD on GPU architecture?


methodA
input

output

Since methodA will be executed many times by different threads on the same input, the logic of methodA must be deterministic to guarantee the correctness in SISD phase

Warp-Centric Programming Method

Advantages:

No path divergence Increase memory coalescing Allows to take advantage of shared memory

Warp-Centric Programming Method

Warp-Centric Programming Method

Traditional approach: thread-to-data


mapping Threads 0 1 2 3 4 5 . . . . n-2 n-1 n

Level[ ] 0 1 2 3 4 5 . . . . n-2 n-1 n

Warp-Centric Programming Method

Warp-centric approach: warp-to-chunk of data


Chunks 0 Warps Threads 0 0 1 2 3 1 1 4 5 . . . . n-2 k k n-1 n

Level[ ] 0 1 2 3 4 5 . . . . n-2 n-1 n

mapping

Warp-Centric Programming Method

Warp-centric approach: warp-to-chunk of data

Coarse-grained mapping Each chunk must be mapped to a warp Number of chunks Number of warps Chunk size & warp size are independent Chunk size is limited by the size of shared memory

Warp-Centric Programming Method

Disadvantages:

If native SIMT width of the user application is small, the underlying hardware will be under-utilized (# threads assigned to data < #physical cores)

Warp-Centric Programming Method

Disadvantages:

The ratio of the SIMT phase duration to the SISD phase duration imposes an Amdahls limit on performance
Amdahls Law:

Improvement:

Warp-Centric Programming Method

Addressing these issues:

Partition a warp into multiple virtual warps with smaller size


Virtual warp size =

Increase parallelism within the SISD phase Improve ALU utilization by O(K) times Drawback: might re-introduce path divergence among these multiple virtual warps. This is the trade-off between path divergence and ALU utilization.

Other Techniques

Deferring Outliners:

Define some threshold Defer processing any node having degree greater than the threshold Process outliners in a separate kernel method

Dynamic Workload Distribution:

Virtual warp-centric does not prevent work imbalance among warps in a block Each warp fetches a chunk of work from shared work queue Trade-off between static/dynamic chunk of work:

static work distribution suffers from work imbalance dynamic work distribution imposes overhead

Experimental Results

Input Graphs:

RMAT: scale-free graph which follows a power law degree distribution like many real world graph. The average vertex degree is 12. RANDOM: uniformly distributed graph instance created by randomly connecting m pairs of nodes out of total n nodes. The average vertex degree is 12. LiveJournal: real world graph, is a very irregular structure Patent: is relatively regular graph, has a smaller average degree

Experimental Results

Experimental Results

Study of Architectural Effects

Advantages of GPU Architecture


Enables massively parallel execution Uses large number of warps to hide memory latency Uses GDDR3 memory which has higher bandwidth and lower latency than FB-DIMM based CPU main memory

Study of Architectural Effects

Effect of bandwidth utilization and latency hiding in GPUs

Conclusion:

Graph algorithms are bound by memory bandwidth

Last But Not Least

A supercomputer is a device for turning compute-bound problems into I/O-bound problems Ken Batcher

You might also like