1

Sockets Direct Protocol Over InfiniBand in Clusters: Is it Beneficial?

P. Balaji, S. Narravula, K. Vaidyanathan, S. Krishnamoorthy, J. Wu and D. K. Panda

Network Based Computing Laboratory

The Ohio State University

Presentation Layout

� Introduction and Background

� Sockets Direct Protocol (SDP)

� Multi-Tier Data-Centers

� Parallel Virtual File System (PVFS)

� Experimental Evaluation

� Conclusions and Future Work

2

Introduction

• Advent of High Performance Networks– Ex: InfiniBand, Myrinet, 10-Gigabit Ethernet

– High Performance Protocols: VAPI / IBAL, GM, EMP

– Good to build new applications

– Not so beneficial for existing applications

• Built around Portability: Should run on all platforms

• TCP/IP based Sockets: A popular choice

• Performance of Application depends on the Performance of Sockets

• Several GENERIC optimizations for sockets to provide high performance

– Jacobson Optimization: Integrated Checksum-Copy [Jacob89]

– Header Prediction for Single Stream data transfer

[Jacob89]: “ An analysis of TCP Processing Overhead” , D. Clark, V. Jacobson, J. Romkey and H. Salwen. IEEE Communications

Network Specific Optimizations

• Generic Optimizations Insufficient– Unable to saturate high performance networks

• Sockets can utilize some network features– Interrupt Coalescing (can be considered generic)

– Checksum Offload (TCP stack has to modified)

– Insufficient!

• Can we do better?– High Performance Sockets

– TCP Offload Engines (TOE)

3

High Performance Sockets

High Performance Network

Pseudo sockets layer

Application or Library

Hardware

Kernel

User Space

SocketsOS Agent

NetworkNative Protocol

NIC

IP

TCP

Sockets

Application or Library

Hardware

Kernel

User Space

Traditional Berkeley Sockets High Performance Sockets

InfiniBand Architecture Overview

• Industry Standard

• Interconnect for connecting compute and I/O nodes

• Provides High Performance

– Low latency of lesser than 5us

– Over 840MBps uni-directional bandwidth

– Provides one-sided communication (RDMA, Remote Atomics)

• Becoming increasingly popular

4

Sockets Direct Protocol (SDP*)• IBA Specific Protocol for Data-Streaming

• Defined to serve two purposes:– Maintain compatibility for existing applications

– Deliver the high performance of IBA to the applications

• Two approaches for data transfer: Copy-based and Z-Copy

• Z-Copy specifies Source-Avail and Sink-Avail messages– Source-Avail allows destination to RDMA Read from source

– Sink-Avail allows source to RDMA Write to the destination

• Current implementation limitations:– Only supports the Copy-based implementation

– Does not support Source-Avail and Sink-Avail

*SDP implementation from the Voltaire Software Stack

Presentation Layout

� Introduction and Background

� Sockets Direct Protocol (SDP)

� Multi-Tier Data-Centers

� Parallel Virtual File System (PVFS)

� Experimental Evaluation

� Conclusions and Future Work

5

Multi-Tier Data-Centers

• Client Requests come over the WAN (TCP based + Ethernet Connectivity)

• Traditional TCP based requests are forwarded to the inner tiers

• Performance is limited due to TCP

• Can we use SDP to improve the data-center performance?

• SDP is not compatible with traditional sockets: Requires TCP termination!

(Courtesy Mellanox Corporation)

3-Tier Data-Center Test-bed at OSU

Database ServersClients

Application Servers

Web Servers

Proxy Nodes

Tier 0

Tier 1

Tier 2

Generate requests for both web servers and

database servers

TCP TerminationLoad Balancing

Caching

Caching

Dynamic Content Caching

Persistent Connections

File System evaluation

Caching Schemes

������

� � �

�

� ��

�� �

������

WAN

6

Presentation Layout

� Introduction and Background

� Sockets Direct Protocol (SDP)

� Multi-Tier Data-Centers

� Parallel Virtual File System (PVFS)

� Experimental Evaluation

� Conclusions and Future Work

Network

Parallel Virtual File System (PVFS)

ComputeNode

ComputeNode

ComputeNode

ComputeNode

Meta-DataManager

I/O ServerNode

I/O ServerNode

I/O ServerNode

MetaData

Data

Data

Data

• Relies on Striping of data across different nodes

• Tries to aggregate I/O bandwidth from multiple nodes

• Utilizes the local file system on the I/O Server nodes

7

Parallel I/O in Clusters via PVFS

• PVFS: Parallel Virtual File System– Parallel: stripe/access data across multiple nodes

– Virtual: exists only as a set of user-space daemons

– File system: common file access methods (open, read/write)

• Designed by ANL and Clemson

���

����������������� �

���

����������������� �

� � �

� ��� � �

���� � �!"!#

��$�%��

����������&�

���� � �!"!#

��$�%��

����������&��

'�&� ��� ���

“ PVFS over InfiniBand: Design and Performance Evaluation” , Jiesheng Wu, Pete Wyckoff and D. K. Panda. International Conference on Parallel Processing (ICPP), 2003.

Presentation Layout

� Introduction and Background

� Sockets Direct Protocol (SDP)

� Multi-Tier Data-Centers

� Parallel Virtual File System (PVFS)

� Experimental Evaluation� Micro-Benchmark Evaluation

� Data-Center Performance

� PVFS Performance

� Conclusions and Future Work

8

Experimental Test-bed

• Eight Dual 2.4GHz Xeon processor nodes

• 64-bit 133MHz PCI-X interfaces

• 512KB L2-Cache and 400MHz Front Side Bus

• Mellanox InfiniHost MT23108 Dual Port 4x HCAs

• MT43132 eight 4x port Switch

• SDK version 0.2.0

• Firmware version 1.17

Latency and Bandwidth Comparison

• SDP achieves 500MBps bandwidth compared to 180MBps of IPoIB• Latency of 27us compared to 31us of IPoIB

• Improved CPU Utilization

Latency and CPU utilization on SDP vs IPoIB

0

10

20

30

40

50

60

70

2 4 8 16 32 64 128

256

512 1K 2K 4K

Message Size

Tim

e (u

s)

0

10

20

30

40

50

60

% C

PU

util

izat

ion

IPoIB CPU SDP CPU IPoIB

SDP VAPI send/recv VAPI RDMA write

Bandwidth and CPU utilization on SDP vs IPoIB

0100200300400500600700800

900

4 16 64 256 1K 4K 16K 64KMessage Size

Ban

dw

idth

(Mb

ytes

/s)

0

40

80

120

160

200

% C

PU

util

izat

ion

IPoIB CPU SDP CPU IPoIBSDP VAPI send/recv VAPI RDMA write

9

Hotspot Latency

Hotspot Latency on SDP vs IPoIB

0

100

200

300

400

500

600

700

1 2 3 4 5 6 7Number of Nodes

Tim

e (u

s)

0

20

40

60

80

100

120140

160

180

200

% C

PU

Util

izat

ion

16K IPoIBCPU 16K SDP CPU 1K IPoIB 1K SDP4K IPoIB 4K SDP 16K IPoIB 16K SDP

• SDP is more scalable in hot-spot scenarios

Data-Center Response TimeClient Response Time

0

50

100

150

200

250

32K

64K

128

K

256

K

512

K

1M

2M

Message Size (bytes)

Res

pons

e T

ime

(ms)

IPoIB SDP

Web Server Delay

0

5

10

15

20

25

32K

64K

128k

256k

512k

1024

k

2048

k

Message Size (bytes)

Tim

e S

pent

(ms)

IPoIB SDP

• SDP shows very little improvement: Client network (Fast Ethernet) becomes the bottleneck

• Client network bottleneck reflected in the web server delay: up to 3 times improvement with SDP

10

Data-Center Response Time (Fast Clients)

0

5

10

15

20

25

30

32K 64K 128K 256K 512K 1M 2M

Message Size (bytes)

Re

spo

nse

Tim

e (m

s)

IPoIBSDP

• SDP performs well for large files; not very well for small files

Data-Center Response Time Split-up

Init + Qtime8% Request Read

3%

Core Processing10%

URL Manipulation1%

Back-end Connect32%Request Write

2%

Reply Read14%

Cache Update2%

Response Write25%

Proxy End3%

Init + Qtime9% Request Read

3%

Core Processing12%

URL Manipulation1%

Back-end Connect14%

Request Write2%Reply Read

15%

Cache Update3%

Response Write38%

Proxy End3%

IPoIB SDP

11

Data-Center Response Time without Connection Time Overhead

0

5

10

15

20

25

30

32K 64K 128K 256K 512K 1M 2M

Message Size (bytes)

Re

spo

nse

Tim

e (m

s)

IPoIBSDP

• Without the connection time, SDP would perform well for all file sizes

PVFS Performance using ramfsRead Bandwidth (3 IODs)

0

500

1000

1500

1 2 3 4 5

No. of Clients

Ban

dwid

th (M

Bps

)

IPoIB SDP VAPI

Write Bandwidth (3IODs)

0200400600

80010001200

1 2 3 4 5No. of Clients

Ban

dwid

th (M

Bps

)

Read Bandwidth (4IODs)

0

500

1000

1500

2000

1 2 3 4No. of Clients

Ban

dwid

th (M

Bps

)

Write Bandwidth (4IODs)

0

500

1000

1500

1 2 3 4No. of Clients

Ban

dwid

th (

MB

ps)

12

PVFS Performance with sync (ext3fs)

60

65

70

75

IPoIB SDP VAPI

Agg

raga

te B

andw

idth

(M

byte

s/s)

• Clients can push data faster to IODs using SDP; de-stage bandwidth remains the same

Conclusions

• User-Level Sockets designed with two motives:– Compatibility for existing applications

– High Performance for modern networks

• SDP was proposed recently along similar lines

• Sockets Direct Protocol: Is it Beneficial?– Evaluated it using micro-benchmarks and real applications

• Multi-Tier Data-Centers and PVFS

– Benefits in environments it’s good for

• Communication intensive environments such as PVFS

– Demonstrate environments it’s yet to mature for

• Connection overhead involving environments such as Data-Centers

13

Future Work

• Connection Time bottleneck in SDP

– Using dynamic registered buffer pools, FMR techniques, etc

– Using QP pools

• Power-Law Networks

• Other applications: Streaming and Transaction

• Comparison with other high performance sockets

�������������������� ��������������

��� ��������������������������

���������������� ������ ������!�

"���#����$����%��������!

Thank You!

NBC Home Page

14

Backup Slides

TCP Termination in SDP

OS by pass

Ethernet

IP

TCP

Proxy

SDP

Infiniband HWEthernet

IP

TCP

Sockets

Browser

Sockets

SDP

Infiniband HW

Web ServerHTTP HTTP

HTML HTML

Personal Notebook/Computer Blade ServersNetwork Service Tier

Ethernet Communication InfiniBand Communication