PTask: Operating System Abstractions to Manage

GPUs as Compute DevicesChris Rossbach, Jon Currey, Microsoft Research

Mark Silberstein, TechnionBaishakhi Ray, Emmett Witchel, UT Austin

SOSP October 25, 2011

There are lots of GPUs◦ 3 of top 5 supercomputers use GPUs◦ In all new PCs, smart phones, tablets ◦ Great for gaming and HPC/batch◦ Unusable in other application domains

GPU programming challenges◦ GPU+main memory disjoint◦ Treated as I/O device by OS

Motivation

PTask SOSP 2011 2

There are lots of GPUs◦ 3 of top 5 supercomputers use GPUs ◦ In all new PCs, smart phones tablets ◦ Great for gaming and HPC/batch◦Unusable in other application domains

GPU programing challenges◦ GPU+main memory disjoint◦Treated as I/O device by OS

Motivation

These two things are related: We need OS abstractions

PTask SOSP 2011 3

The case for OS support PTask: Dataflow for GPUs Evaluation Related Work Conclusion

Outline

PTask SOSP 2011 4

Traditional OS-Level abstractions

programmer-visible interface

OS-level abstractionsHardware interface

1:1 correspondence between OS-level and user-level abstractions

PTask SOSP 2011 5

DirectX/CUDA/OpenCL Runtime

Language Integration

Shaders/Kernels

GPGPUAPIs

GPU Abstractions

programmer-visible interface

1 OS-level abstraction!

PTask SOSP 2011 6

1. No kernel-facing API2. No OS resource-management3. Poor composability

CPU-bound processes hurt GPUs

• Image-convolution in CUDA• Windows 7 x64 8GB RAM• Intel Core 2 Quad 2.66GHz• nVidia GeForce GT230

invo

catio

ns p

er s

econ

d

Higher is better no CPU load high CPU load0

200400600800

10001200

GPU benchmark throughput

CPU scheduler and GPU scheduler not integrated!

PTask SOSP 2011 7

GPU-bound processes hurt CPUs

• Windows 7 x64 8GB RAM• Intel Core 2 Quad 2.66GHz• nVidia GeForce GT230

Flatter lines Are better

OS cannot prioritize cursor updates• WDDM + DWM + CUDA == dysfunction

PTask SOSP 2011 8

Composition: Gestural Interface

capture

filterxform

“Hand” events

Raw images

detect

High data rates Data-parallel algorithms

… good fit for GPU

PTask SOSP 2011 9

noisy point cloud

geometric transformation

capture camera images

detect gestures

noise filtering

NOT Kinect: this is a harder problem!

What We’d Like To Do#> capture | xform | filter | detect &

PTask SOSP 2011 10

Modular design flexibility, reuse

Utilize heterogeneous hardware Data-parallel components GPU Sequential components CPU

Using OS provided tools processes, pipes

CPU CPUGPU GPU

GPUs cannot run OS: different ISA Disjoint memory space, no coherence Host CPU must manage GPU execution

◦ Program inputs explicitly transferred/bound at runtime◦ Device buffers pre-allocated

GPU Execution model

CPUMain memory

GPU memory

GPU

Copy inputs Copy outputs Send commands

User-mode appsmust implement

PTask SOSP 2011 11

OS executive

capture

GPU

Data migrationke

rnel

user

Run!

camdrv GPU driver

HW

PCI-xfer PCI-xfer

xform

copyto

GPU

copyfrom GPU

PCI-xfer PCI-xfer

filter

copyfrom GPU

detect

IRP

HIDdrv

read()

copyto

GPU

write() read() write()read() write() read()

capture xform filter detect

#> capture | xform | filter | detect &

PTask SOSP 2011 12

GPU Analogues for:◦ Process API◦ IPC API◦ Scheduler hints

Abstractions that enable:◦ Fairness/isolation◦ OS use of GPU◦ Composition/data movement optimization

GPUs need better OS abstractions

PTask SOSP 2011 13

The case for OS support PTask: Dataflow for GPUs Evaluation Related Work Conclusion

Outline

PTask SOSP 2011 14

ptask (parallel task) ◦ Has priority for fairness◦ Analogous to a process for GPU execution◦ List of input/output resources (e.g. stdin, stdout…)

ports◦ Can be mapped to ptask input/outputs◦ A data source or sink

channels◦ Similar to pipes, connect arbitrary ports◦ Specialize to eliminate double-buffering

graph◦ DAG: connected ptasks, ports, channels

datablocks◦ Memory-space transparent buffers

PTask OS abstractions: dataflow!

• OS objectsOS RM possible• data: specify where, not how

PTask SOSP 2011 15

PTask Graph: Gestural Interface

PTask SOSP 2011 16

xform

ptask (GPU)

port

channel

cloud

rawi

mg filter

f-out

f-incapture

process (CPU)

detect

ptask graph

ptask graph

#> capture | xform | filter | detect &

mapped mem GPU mem GPU mem

Optimized data movement

raw

im g

Data arrival triggers computation

datablock

Graphs scheduled dynamically◦ ptasks queue for dispatch when inputs ready

Queue: dynamic priority order◦ ptask priority user-settable◦ ptask prio normalized to OS prio

Transparently support multiple GPUs◦ Schedule ptasks for input locality

PTask Scheduling

PTask SOSP 2011 17

PTask SOSP 2011 18

Main Memory

GPU 0 Memory

GPU 1 Memory

…Datablock

space V M R

W datamain 1 1 1 1gpu0 0 1 1 0gpu1 1 1 1 0

Logical buffer ◦ backed by multiple physical buffers◦ buffers created/updated lazily◦ mem-mapping used to share across process boundaries

Track buffer validity per memory space◦ writes invalidate other views

Flags for access control/data placement

Location Transparency: Datablocks

Datablockspac

e V M RW data

main 0 0 0 0gpu 0 0 0 0

Main Memory

GPU Memory

Datablock Action Zone

PTask SOSP 2011 19

xform

cloud

rawi

mg filterf-incapture

#> capture | xform | filter …

ptaskport

channel

process

datablock1 1 1 11 1 1 11 1 1 11 1 1 10 1 1 1

raw

im g

clou

d

…

Revised technology stack

• 1-1 correspondence between programmer and OS abstractions• GPU APIs can be built on top of new OS abstractions

datablock

port

port

PTask SOSP 2011 20

The case for OS support PTask: Dataflow for GPUs Evaluation Related Work Conclusion

Outline

PTask SOSP 2011 21

Windows 7◦ Full PTask API implementation◦ Stacked UMDF/KMDF driver

Kernel component: mem-mapping, signaling User component: wraps DirectX, CUDA, OpenCL

◦ syscalls DeviceIoControl() calls Linux 2.6.33.2

◦ Changed OS scheduling to manage GPU GPU accounting added to task_struct

Implementation

PTask SOSP 2011 22

Windows 7, Core2-Quad, GTX580 (EVGA) Implementations

◦ pipes: capture | xform | filter | detect◦ modular: capture+xform+filter+detect, 1process◦ handcode: data movement optimized, 1process◦ ptask: ptask graph

Configurations◦ real-time: driven by cameras◦ unconstrained: driven by in-memory playback

Gestural Interface evaluation

PTask SOSP 2011 23

runtime user sys0

0.5

1

1.5

2

2.5

3

3.5

handcodemodularpipesptask

rela

tive

to

hand

code

Gestural Interface Performance

• Windows 7 x64 8GB RAM• Intel Core 2 Quad 2.66GHz• GTX580 (EVGA)

lower is better compared to pipes

• ~2.7x less CPU usage• 16x higher throughput• ~45% less memory usage

compared to hand-code• 11.6% higher throughput• lower CPU util: no driver

program

PTask SOSP 2011 24

Performance Isolation

2 4 6 80200400600800

1000120014001600

fifopriority

PTask priority

PTas

k in

voca

tions

/sec

ond

• Windows 7 x64 8GB RAM• Intel Core 2 Quad 2.66GHz• GTX580 (EVGA)

Higher is better

• FIFO – queue invocations in arrival order• ptask – aged priority queue w OS priority • graphs: 6x6 matrix multiply• priority same for every PTask node

PTask provides throughput proportional to priority

PTask SOSP 2011 25

ptask

Multi-GPU Scheduling

PTask SOSP 2011 26

depth

-1

depth

-2

depth

-4

depth

-60

0.20.40.60.8

11.21.41.61.8

prioritydata-aware

Spee

dup

over

1 G

PU

• Windows 7 x64 8GB RAM• Intel Core 2 Quad 2.66GHz• 2 x GTX580 (EVGA)

Higher is better

• Data-aware == priority + locality• Graph depth > 1 req. for any benefit

Data-aware provides best throughput, preserves priority

• Synthetic graphs: Varying depths

GPU/CPU

cuda-1Linux

cuda-2Linux

cuda-1PTask

cuda-2 Ptask

Read 1.17x

-10.3x -30.8x 1.16x 1.16x

Write 1.28x

-4.6x -10.3x 1.21x 1.20x

Linux+EncFS Throughput

PTask SOSP 2011 27

R/W bnc

EncFS FUSE libc

Linux 2.6.33

SSD1

cuda-1 cuda-2

SSD2 GPU

user-prgs

user-libs

OS

HW

PTask

…

• EncFS: nice -20• cuda-*: nice +19• AES: XTS chaining• SATA SSD, RAID• seq. R/W 200 MB

Simple GPU usage accounting• Restores performance

The case for OS support PTask: Dataflow for GPUs Evaluation Related Work Conclusion

Outline

PTask SOSP 2011 28

OS support for heterogeneous platforms:◦ Helios [Nightingale 09], BarrelFish [Baumann 09] ,Offcodes

[Weinsberg 08] GPU Scheduling

◦ TimeGraph [Kato 11], Pegasus [Gupta 11] Graph-based programming models

◦ Synthesis [Masselin 89]◦ Monsoon/Id [Arvind]◦ Dryad [Isard 07]◦ StreamIt [Thies 02]◦ DirectShow ◦ TCP Offload [Currid 04]

Tasking◦ Tessellation, Apple GCD, …

Related Work

PTask SOSP 2011 29

OS abstractions for GPUs are critical◦ Enable fairness & priority◦ OS can use the GPU

Dataflow: a good fit abstraction◦ system manages data movement◦ performance benefits significant

Conclusions

Thank you. Questions?

PTask SOSP 2011 30