Cisco - Enabling High Performance Grids and Utility Computing · Cisco - Enabling High Performance ... Traditional Model Kernel Bypass Model Kernel User TCP/IP ... High Performance

1© 2005 Cisco Systems, Inc. All rights reserved.

Cisco - Enabling High Performance Grids and Utility Computing

Shankar SubramanianTechnical Director – Storage & Server NetworkingCisco Systems

222© 2005 Cisco Systems, Inc. All rights reserved.© 2005 Cisco Systems, Inc. All rights reserved.

Agenda

• InfiniBand Hardware & System Overview

• RDMA and Upper Layer Protocols

• High Performance Computing Architectures

• HPC Building-blocks

• I/O Virtualization

• Server Virtualization

InfiniBand Overview



What Is InfiniBand?

• InfiniBand is a high speed – low latency technology used to interconnect servers, storage and networks within the datacenter

• Standards Based – InfiniBand Trade Associationhttp://www.infinibandta.org

• Scalable Interconnect:1X = 2.5Gb/s

4X = 10Gb/s

12X = 30Gb/s


InfiniBand Physics

• Copper and Fiber interfaces are specified

• Copper Up to 15m* for 4x connections

Up to 10m for 12x connections

• OpticalInitial availability via dongle solution

Up to 300m with current silicon

Long Haul possible, but not with current silicon

* 20m in certain circumstances


InfiniBand Physics

• Link is bonded 2.5Gbps (1x) linksFiber is a ribbon cable

Copper is a multi-conductor cable

• Each Link is 8b/10b encoded4x Link is 4 2.5Gbps Physical Connections

Each connection is 2Gbps data

SAR provides a single 8Gbps data connection (4x)

24 Gbps (12x)


Pluggable Optics ModuleTransforms Powered Copper Ports to Optical Ports

Topspin Optical Module

• Coverts a copper port to an optical port on a port by port basis

• Extends port to port reach to 150m - 300m with fibreribbon cables


ServerServer ServerServer

InfiniBand Nomenclature

HostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHost

HostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostHostServerServerServerServer

CPU

CPU

Hos

t Int

erco

nnec

t

MemCntlr

SystemMemory

IB L

ink

HCA

SM

Switch

IB Link TCA

IB Link TCA

Ethernet linkIB Link

FC link

HCA – Host Channel AdaptorSM - Subnet managerTCA – Target Channel Adaptor


InfiniBand Switch Hardware

• Hardware switch devices is a cut-through memory switch

• Full-duplex, non-blocking 24 port tag forwarding switch

• Tags are system Local ID, provided to all network endpoints by the Master Subnet Manager on system startup


InfiniBand Host Channel Adapter

• Network interface for IB attached Servers

• Provides hardware Virtual/Physical memory mapping, Direct Memory Access (DMA), and memory protection

• Provides RDMA (Remote DMA) data transfer engine and reliable packet forwarding capabilities


InfiniBand Gateway

• Technically a Target Channel Adapter

• Similar to an HCA attached to an embedded device

• Usually doesn’t require virtual memory manipulation and mapping

• Simplified HCA on a specialized deviceExamples, Ethernet to InfiniBand or Fibre Channel to InfiniBand packet forwarding engines

RDMA and Upper Layer Protocols

121212© 2005 Cisco Systems, Inc. All rights reserved.© 2005 Cisco Systems, Inc. All rights reserved.DC-210811392_06_2005_X


NIC

Current NIC Architecture

CPU

CPU

Hos

t Int

erco

nnec

tMemCntlr

Server (Host)

inte

rcon

nect

System Memory

OS Buffer

App Buffer

Data traverses bus 3 times

Multiple context switches robs CPU cycles from actual workMemory bandwidth and per packet interrupts limit max throughputOS manages end-to-end communications path


With RDMA and OS Bypass

HCA

CPU

CPU

Hos

t Int

erco

nnec

tMemCntlr

Server (Host)

inte

rcon

nect

System Memory

OS Buffer

App Buffer

Data traverses bus once, saving CPU and memory cycles

Secure Memory – Memory transfers with no CPU overheadPCI-X/PCI-e becomes the bottleneck for network data transmissionHCA manages remote data transmission


Kernel Bypass

Hardware

Application

Traditional Model Kernel Bypass Model

KernelUser

TCP/IPTransport

Driver

Hardware

Application

Kernel

User

TCP/IPTransport

Driver

RDMA ULP

SocketsLayer

SocketsLayer


Upper Layer Protocols

• Variety of software protocols to handle high speed communication over RDMA

• Protocols includeIP-over-InfiniBand – IETF http://www.ietf.org/internet-drafts/draft-ietf-ipoib-ip-over-infiniband-09.txtSDP (Socket Direct Protocol) – InfiniBand Trade Association http://infinibandta.orgSRP (SCSI RDMA Protocol) – ANSI T10 http://www.t10.org/ftp/t10/drafts/srp/srp-r16a.pdfDAPL (Direct Access Provider Library) – DAT Collaborative http://www.datcollaborative.orgMPI (Message Passing Interface) – MPI Forum http://www.mpi-forum.org


InfiniBand PerformanceMeasured Results

BSD Sockets Async I/O extension

Application

1GE

DirectAccess

IPoIB

TCPIP

SDP

10G IB

uDAPLSRP MPI

3.5 usec8 usec18 usec18 usec30 usec40-60 usecLatency

8Gb/s8Gb/s7.9Gb/s4.5 Gb/s4.1Gb/s1 Gb/sThroughput

High Performance Computing



High Performance Computing Applications

• Parallel processing applicationsClosely coupled

Finite Element Analysis (Crash Simulation)

Fluid Dynamics (Injection Molding)

Loosely coupled

Dataset searches (Terabyte->Petabyte datasets)

Monte-Carlo simulation (10,000s of repetitions)


High Performance Computing Networks

• Two Standards Based TechnologiesGigabit Ethernet/10 Gigabit EthernetInfiniBand

• Multiple UsesHPC interconnectStorage trafficLoad/Unload data movementApplication/Systems management


Network Types

• Network style is guided by applicationClosely coupled applications

Latency is a problem

Throughput is key to resolving latency issues

Loosely coupled applications

Load/unload of datasets can be a key bottleneck

Low latency for rapid response is critical


IB Routing for HPC

• Subnet Manager provides global route engine for entire IB Fabric

• Shortest Path First routing

• Round Robin load balancing

• Static routes


IB Routing for HPC


HPC Storage Problem

• As the job grows so does the storageClusters growing to 1000 nodes or more

Storage growing into the PetaByte range

• The storage problemLoad/Unload a large dataset

Get shared access to large datasets on the fly

Performance is an issue: Need multi-GigaByte/sthroughput


Current Solutions

• NFSOld standby: Doesn’t scale well (single server)

• Current commercial cluster file systemsDesigned for multiple reader/writer situations

Don’t scale beyond 10s of nodes

Don’t necessarily manage the throughput problem

Need to build out separate Fibre Channel fabric


HPC Storage Solutions

• Next Generation Cluster File-systemsBased on new file-systems or modification of old (nfs or iSCSI)

Split the data across multiple file service hosts

Either act as a RAIF (Redundant Array of Inexpensive File-servers), or allow clients to access any one file server, while everyone shares the same back-end storage devices


InfiniBand Storage Acceleration

• Further accelerate storage accessIB based forwarding over SDP or DAPL can provide CPU offload and increased throughput

Low latency improves overall throughput and performance

High bandwidth improves load/unload times dramatically


HPC Glossary

• Fat-tree – Non-blocking switch architecture (akaCLOS)

• Bisectional bandwidth – the total system bandwidth across the middle of the network

• Non-blocking – full host bandwidth all-2-all communication

A New Infrastructure Category



Network Switch

Clients

Network Resources (Internet, Printer, Server)

Storage Switch

Server

Storage (SAN)

Server Switch

Servers

StorageNetwork

A New Category of Data Center Infrastructure-The Server Fabric Switch


Virtualization (I/O, Storage, and CPU)

−Shared Resources Across Entire Cluster−Routing, Aggregation, Load Balancing−App/OS to CPU provisioning

High Performance Server-to-Server

Interconnect

−RDMA−High Bandwidth −Low Latency−InfiniBand today; PCI-Express and /or 10GigE when ready

Policy-Based Dynamic Resource Mapping

What Makes The Server Fabric Switch Different?

Performance Performance andand ControlControl


Server Fabric Switch ApplicationsWhy Performance and Control?

Server Clustering

−High Performance Computing (HPC)−“Enterprise-Class” HPC −Database Scalability

Utility or Grid Computing

−Application Provisioning−Server Re-purposing−Server Migration

Applications

I/O Virtualization

−I/O Consolidation−I/O Aggregation−Server Consolidation


Cisco Building Blocks (Server Fabric Switching)

Gateway Modules- InfiniBand to Ethernet- InfiniBand to Fibre Channel

Integrated System and Fabric management

Switches

Host Channel Adapter (HCA)With upper layer protocols

SRPSDPuDAPLMPIIPoIB

Linux and Windows driver support


Serv

er F

abric

Sw

itch

HC

A

The Cisco SFS Product Line

•(2) 4XIB PCI-X

•(2) 4XIB PCI-ex

Bla

de

Serv

er

•Remote Boot

•Linux Host Driver

•Windows Host Driver

IBM BladeCenter

• HCA (2) 1XIB PCI-X

•Embedded switch (14) 1XIB (Internal) + (1) 4XIB and (1) 12XIB (External)

Dell 1855

• HCA (2) 4XIB PCI-ex

• Passthru Module (10) 4XIB

Infin

iBan

dM

ultif

abric

SFS 7000 (TS120)

(24) 4XIB(24) 4XIB

SFS 7008 (TS270)

(96) 4XIB(96) 4XIB

SFS 3012 (TS360)

(24) 4XIB + 12 (24) 4XIB + 12 GwsGws

SFS 3001 (TS90)

(12) 4XIB + 1 (12) 4XIB + 1 GwGw

•(2) 2G FC GW

Softw

are

VFrame™ Server Fabric

Virtualization Software R3.0

•(6) GE GW

*plus InfiniBand Cables

I/O Virtualization



The Evolution of I/O Virtualization

Pro: Single managed entity, fast backplane

Con: Expensive, Proprietary server +

backplane

SMP

Pro: Standard servers, inexpensive

Con: Lots of managed components, low-

performing interconnect

Dis-aggregation

Ethernet N

etwork

Fibre Channel

Virtualization

Pro: Reduced # of managed

components, virtual I/O, fast standards

backplane


Evolution of the Data CenterNetwork and Storage Virtualization


Evolution of the Data CenterServer Virtualization - The Server Switch


Virtual I/O for Network and StorageUnified “wire-once” fabric

SAN CiscoSFS 3012

LAN/WAN

Server Cluster

Fibre Channel to InfiniBand gateway for storage access

Two 2-Gbps Fibre Channel ports per gatewayCreate 10-Gbps virtual storage pipe to each server

Fibre Channel to InfiniBand gateway for storage access

Two 2-Gbps Fibre Channel ports per gatewayCreate 10-Gbps virtual storage pipe to each server

Ethernet to InfiniBand gateway for LAN access

Six Gigabit Ethernet ports per gatewayCreate virtual GigE pipe to each server

Ethernet to InfiniBand gateway for LAN access

Six Gigabit Ethernet ports per gatewayCreate virtual GigE pipe to each server

Single InfiniBand link for:- Storage- Network

Single InfiniBand link for:- Storage- Network


Virtual I/O for Blade ServersEliminating I/O Bottlenecks

SAN Server Switch LAN/WAN

Server Blades

Integrated 10Gbps InfiniBand switch or pass-thru module

Integrated 10Gbps InfiniBand switch or pass-thru module

Shared Ethernet I/O across blades and racksShared Ethernet I/O across blades and racks

Shared FC I/O across blades and racksShared FC I/O across blades and racks


Integrated InfiniBand for Blade Servers Create “wire-once” fabric

• Integrated 10Gbps InfiniBand switches provide unified “wire-once” fabric

• Optimize density, cooling, space, and cable management.

• Virtual I/O provides shared Ethernet and Fibre Channel ports across blades and racks

• Option of integrated InfiniBand switch (ex: IBM BC) or pass-thru module (ex: Dell 1855)

IB SwitchIB Switch

10Gbps 30Gbps

Blade Chassis with InfiniBand Switches

HCA

Virtual I/O: How it Works



Transparent Topology Architecture

• IP CommunicationsInside Fabric: IP over InfiniBand (IPoIB) enables transparent communications for any IP-based applications. Outside Fabric: InfiniBand-to-Ethernet Gateways provide transparent access to existing IP Ethernet infrastructure.

• Fibre Channel StorageInside Fabric: SCSI RDMA Protocol (SRP) enables SCSI over InfiniBand.Outside Fabric: InfiniBand-to-Fibre Channel Gateways provide transparent access to existing SAN infrastructure.

• Cluster CommunicationsInside Fabric: Remote DMA protocols for transparent performance enhancements.


The InfiniBand Driver Architecture

BSD Sockets FS API

TCPSDP

IP

DriversVAPI

ETHER INFINIBAND HCA

DAT FILE SYSTEM

SCSI

SRP

FC

FCP

SDP

INFINIBAND SAN

TS API

BSD Sockets NFS-RDMA

LAN/WAN UNIFIED FABRICSAN

INFINIBAND SWITCHETHERSWITCH

FCSWITCHFC GW

EETH GW

NETWORK

APPLICATION

UDAPL

TS TS

IPoIB

User

Kernel


InfiniBand-to-Ethernet Gateway Overview

• Ensures seamless integration with IP-based applications.

• Act like L2 bridge between IB and Ethernet

• Bridge group is the main forwarding entity

• Bridge group has two bridge ports Ethernet and IPoIB

• Bridge group bridges one VLAN to one IB partition

• Ethernet bridge port can be tagged or untagged

• Ethernet bridge port can aggregate up to 6 ports


InfiniBand-to-Ethernet Gateway Features

• IP-Only protocols

• 802.1Q VLAN support

• Link aggregation

• IPv4 multicast support

• Loop protection

• Ethernet jumbo frames up to 9k

• IP fragmentation

• High availability


InfiniBand-to-Fibre Channel Gateway

• Ensures seamless integration with important SAN tools.Fabric-based Zoning

LUN-based access controls

Storage and host-based HA and load balancing tools

• Creates SAN network addresses on InfiniBand.SAN Management Tools must “see” each node.

Creates “talk-through” mode with virtual WWNNs per server.

• Enables SAN Interoperability with InfiniBand.Appears as virtual NL-Port, N-Port, E-Port.

Proven interoperability with Cisco MDS, Brocade, McData, Qlogic, EMC, IBM, Hitachi, and more.


Typical SAN Today

LUN0

LUNn

LUN0

LUNn

LUN0

LUNn

FC RAIDComplex

WWNN_a

WWNN_c

WWNN_b

LUN0

LUNn

LUN0

LUNn

LUN0

LUNn

FC RAIDComplex

WWNN_d

WWNN_f

WWNN_e

IDC

FibreChannelFabric

Sni ffe r Serverm on itoring/analys is



Assumptions Made by SAN Mgmt Tools:

Each Physical Connection is a Logical Connection

Each Physical Connection Has a Unique Address

Logical Connectivity Access Between Nodes Can be Controlled by Physical Address


Sniffer Servermonitoring/analysis





Physical vs. Logical View






Physical View

Logical View

Each server assignedunique WWNN.Appears as direct-attached N_Port.

Server Virtualization



Three Categories of Server Virtualization

• Virtual Machine: Splits a servers into independent virtual servers.

VMWare, XEN, MSFTMain value is higher server utilization.

• Virtual SMP: Combines servers together into a single managed powered entity.

Virtual Iron, QlustersMain value is scaling mission critical apps on commodity HW.

• Physical Server Virtualization: Makes servers stateless by moving server identity into the network, including storage and I/O subsystem.

Cisco VFrame™, EgeneraMain value is making infrastructure change easier in heterogeneous environment.


VFrame Server Virtualization FrameworkBuilding Blocks

Ethernet (I/O)

InfiniBand (RDMA)High Performance

Server-Server Connectivity

Fibre Channel (Storage)

Topology Transparency

Virtualization and Boot Services

Policy and Provisioning Services


VFrame™

• Software suite that makes the Server Switch programmable• Three main components

VFrame™ Embedded System LogicPolicy ingestion, interpretation, and enforcement at the server switch

VFrame™ APIs (and SDK) Allows 3rd party (End-user Customers, Software Partners, System Vendor OEMs) management and provision tools to program and manage the server switch fabric Software Partners

VFrame™ Director Software package disseminates policies to server switch fabric Central policy enforcement provides better system wide decision making and conflict arbitrationCan be installed on any server in the network


ProgrammabilityVFrame™

1) Server Switch receives policy from VFrame™ Director or 3rd party software.

2) Based on policy, Server Switch assembles the virtual server

Selects server(s) that meet minimum criteria (e.g. CPU, memory)Boot server(s) over the network with appropriate app/os imageCreates virtual IPs in servers and maps to VLANs for client access.Creates virtual HBAs in servers and maps to Zones, LUNs, and WWNNs for storage access

Policy

Virtual Server

SAN

LAN

CPUs

vIPvIP

vHBAvHBA

vHBA

Cisco SFS 3012


Case Study: Large Wall Street BankEnterprise Grid Computing

• Application: Replace proprietary platforms with standards-based componentsBuild scalable “on-demand” compute grid for financial applications

• Benefits: 20X Price/Performance Improvement over four years30-50% Application Performance ImprovementStandards-based solution for on-demand computingEnvironment that scales using 500-node building blocksCentralized shared I/O pool for flexibly allocating SAN/IP bandwidth

LAN

Storage

ExistingN/W

44x 24-port

GridI/O

CoreFabric

Edge

12 hosts512 Nodes


Additional Content

• Cisco Press –

http://www.cisco.com/go/datacenter

• Cisco Press –

http://ciscopress.com/datacenterfundamentals

Q and A



Cisco - Enabling High Performance Grids and Utility Computing · Cisco - Enabling High Performance ... Traditional Model Kernel Bypass Model Kernel User TCP/IP ... High Performance

Documents