KAIST Computer Architecture Lab. The Effect of Multi-core on HPC Applications in Virtualized Systems Jaeung Han¹, Jeongseob Ahn¹, Changdae Kim¹, Youngjin.

KAISTComputer Architecture Lab.

The Effect of Multi-core on HPC Applica-tions in Virtualized Systems

Jaeung Han¹, Jeongseob Ahn¹, Changdae Kim¹, Youngjin Kwon¹, Young-ri Choi², and Jaehyuk Huh¹

¹ KAIST(Korea Advanced Institute of Science and Technology)

² KISTI(Korea Institute of Science and Technology Information)

Outline

• Virtualization for HPC

• Virtualization on Multi-core

• Virtualization for HPC on Multi-core

• Methodology

• PARSEC – shared memory model

• NPB – MPI model

• Conclusion

2

Outline




• Methodology



• Conclusion

3

Benefits of Virtualization

4

Hardware

Virtual Machine Monitor

VM VM VM

• Improve system utilization by consolidation


5

Hardware


VMWin-dows

VM

Linux

VM

Solaris

• Improve system utilization by consolidation• Support for multiple types of OSes on a system


6

Hardware


VMWin-dows

VM

Linux

VM

Solaris

• Improve system utilization by consolidation• Support for multiple types of OSes on a system• Fault isolation


7

Hardware


VMWin-dows

VM

Linux

VM

Solaris

Hardware


• Improve system utilization by consolidation• Support for multiple types of OSes on a system• Fault isolation• Flexible resource management


8

• Improve system utilization by consolidation• Support for multiple types of OSes on a system• Fault isolation• Flexible resource management

Hardware


VMWin-dows

VM

Linux

VM

Solaris

Hardware



9

• Improve system utilization by consolidation• Support for multiple types of OSes on a system• Fault isolation• Flexible resource management• Cloud computing

VMWin-dows

VM

Linux

VM

Solaris Cloud

Hardware


Virtualization for HPC

• Benefits of virtualization

– Improve system utilization by consolidation

– Support for multiple types of OSes on a system

– Fault isolation

– Flexible resource management

– Cloud computing

• HPC is performance-sensitive

• Virtualization can help HPC workloads

10

resource-sensitive

Outline




• Methodology



• Conclusion

11

Virtualization on Multi-core

12

core

• More VMs on a physical machine• More complex memory hierarchy (NUCA, NUMA)

VM

VM

core

VM

VM

core

VM

VM

core

VM

VM

core

VM

VM

core

VM

VM

Shared cache Shared cache

Memory Memory

core

VM

VM

core

VM

VM

Challenges

• VM management cost • Semantic gaps– vCPU scheduling, NUMA

13


VM

VM

VM

VM

VM

VM

VM

VM

Scheduling, Mem-ory, Communica-

tion,I/O multiplexing…

Mem

Mem

core

core

core

core

core

core

core

core


core

core

core

core

OS

Memory

$ $

Outline




• Methodology



• Conclusion

14

Virtualization for HPC on Multi-core

• Virtualization may help HPC• Virtualization on multi-core may have some overheads• For servers, improving system utilization is a key factor• For HPC, performance is a key factor.

15

How much overheads are there?

Where do they come from?

Outline




• Methodology



• Conclusion

16

Machines

• Single Socket System– 12-cores AMD processor– Uniform memory access la-

tency– Two 6MB L3 caches shared

by 6 cores

• Dual Socket System – 2x 4-core Intel processor– Non-uniform memory ac-

cess latency– Two 8MB L3 caches shared

by 4 cores

17

P

L2

P

L2

L3

P

L2

P

L2P

L2

P

L2

P

L2

P

L2

L3

P

L2

P

L2P

L2

P

L2

Single socket: 12-core CPU

Memory

P

L2

P

L2

P

L2

P

L2

L3

P

L2

P

L2

P

L2

P

L2

L3

Dual socket: 2x 4-core CPUs

Workloads

• PARSEC– Shared memory model– Input: native– On one machine

• Single and Dual socket

– Fix: One VM– Vary: 1, 4, 8 vCPUs

• NAS Parallel Benchmark– MPI model– Input: class C– On two machines (dual socket)

• 1Gb Ethernet switch

– Fix: 16 vCPUs– Vary: 2 ~ 16 VMs

18

Mem

Mem

core

core

core

core

core

core

core

core


core

core

core

core

OS

Memory

$ $

Virtual Machine Moni-tor

VM

VM

VM

VM

VM

VM

VM

VM

Hardware

Virtual Machine Moni-tor

VM

VM

VM

VM

VM

VM

VM

VM

Hardware

Semantic gaps VM management cost

Outline




• Methodology



• Conclusion

19

PARSEC – Single Socket

• Single socket• No NUMA effect• Very low virtualization overheads

20

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.81 vCPU4 vCPUs8 vCPUs

2~4 %

Execution times normalized to native runs

PARSEC – Single Socket

• Single socket + pin vCPU to each pCPU• Reduce semantic gaps by prevent vCPU migration• vCPU migration has negligible effect

21

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6



Similar to un-pinned

PARSEC – Dual Socket

• Dual socket, unpinned vCPUs• NUMA effect semantic gap• Significant increase of overheads

22

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8 1 vCPU4 vCPUs8 vCPUs

16~37 %


PARSEC – Dual Socket

• Dual socket, pinned vCPUs• May reduce NUMA effect also• Reduced overheads with 1 and 4 vCPUs

23

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6



XEN and NUMA machine

• Memory allocation policy– Allocate up to 4GB chunk on

one socket

• Scheduling policy– Pinning to allocated socket– Nothing more

• Pinning 1 ~ 4 vCPUs on the socket mem. allocated is possible

• Impossible with 8 vCPUs

24

Mem

core

core

core

core

core

core

core

core

$ $

Mem

VM

0VM

1

VM

2VM

3

Mitigating NUMA Effects

• Range pinning

– Pin vCPUs of a VM on a socket

– Work only if # of vCPUs < # of cores on a socket

– Range-pinned (best): memory of VM in the same socket

– Range-pinned (worst): memory of VM in the other socket

• NUMA-first scheduler

– If there is an idle core in the socket memory allocated, pick it

– If not, anyway, pick a core in the machine

– All vCPUs are not active all the time (sync. or I/O)

25

Range Pinning

• For 4 vCPUs case• Range-pinned(best) ≈ Pinned

26

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Unpinned

Range-pinned (worst)

Range-pinned (best)

Pinned


NUMA-first Scheduler

• For 8 vCPUs case• Significant improvement by NUMA-first scheduler

27

blacksc

holes

cannea

lfer

ret

fluid

anim

ate

freqm

ine

strea

mclu

ster

swap

tions

x264AVG

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8Unpinned

Pinned

NUMA-first


Outline




• Methodology



• Conclusion

28

VM Granularity for MPI model

• Fine-grained VMs– Few processes in a VM– Small VM: vCPUs, memory– Fault isolation among pro-

cesses in different VMs– Many VMs on a machine– MPI communications

mostly through the VMM

•

• Coarse-grained VMs– Many processes in a VM– Large VM: vCPUs, memory– Single failure point for pro-

cesses in a VM– Few VMs on a machine– MPI communications

mostly within a VM

29

VMM

Hardware

VMM

Hardware

VMM

Hardware

VMM

Hardware

NPB - VM Granularity• Work to do are same for all granularity• 2 VMs: each VM has 8 vCPUs, 8 MPI processes• 16 VMs: each VM has 1 vCPU, 1 MPI processes

30

BT CG EP FT IS LU MG SP AVG0

0.5

1

1.5

2

2.5

3 2 VMs4 VMs8 VMs16 VMs


11~54 %

NPB - VM Granularity

• Fine-grained VMs significant overheads (avg. 54%)

– MPI communications mostly through VMM

• Worst in CG with high communication ratio

– Small memory per VM

– VM management costs of VMM

• Coarse-grained VMs much less overheads (avg. 11%)

– Still dual socket, but less overheads than shared memory model

the bottle neck is moved to communication

– MPI communication largely within VM

– Large memory per VM

31

Outline




• Methodology



• Conclusion

32

Conclusion

• Questions on virtualization for HPC on multi-core system– How much overheads are there?– Where do they come from?

• For shared memory model– Without NUMA little overheads– With NUMA large overheads from semantic gaps

• For MPI model– Less NUMA effect communication is important– Fine-grained VMs have large overheads

• Communication mostly through VMM• Small memory / VM management cost

• Future Works– NUMA-aware VMM scheduler– Optimize communication among VMs in a machine

33

34

Thank you!

35

Backup slides

PARSEC CPU Usage

• Environments: native linux, turn on only 8 cores (use 8 threads mode)

• Get CPU usage every seconds, then average them

• For all workloads, less than 800% (fully parallel) NUMA-first can work

36

blackscholes canneal ferret fluidanimate freqmine streamcluster swaptions x264 Avg.0.00%

100.00%

200.00%

300.00%

400.00%

500.00%

600.00%

700.00%

800.00%

KAIST Computer Architecture Lab. The Effect of Multi-core on HPC Applications in Virtualized Systems Jaeung Han¹, Jeongseob Ahn¹, Changdae Kim¹, Youngjin.

Documents

multicore virtualization

hpc virtualization

system slide

hpc benefits of virtualization

outline virtualization

core cpus slide

effect of multicore

consolidation slide