Roland Dreier Technical Lead – Cisco Systems, Inc. OpenIB Maintainer Sean Hefty Software Engineer...
-
Upload
jesse-cunningham -
Category
Documents
-
view
224 -
download
0
Transcript of Roland Dreier Technical Lead – Cisco Systems, Inc. OpenIB Maintainer Sean Hefty Software Engineer...
Roland DreierRoland DreierTechnical Lead – Cisco Systems, Inc.Technical Lead – Cisco Systems, Inc.
OpenIB MaintainerOpenIB Maintainer
Sean HeftySean HeftySoftware Engineer – Intel CorporationSoftware Engineer – Intel Corporation
OpenIB MaintainerOpenIB Maintainer
Yaron HavivYaron HavivCTO – VoltaireCTO – Voltaire
Stack Development Stack Development Update, Testing Update, Testing Status and Target Status and Target Release GoalsRelease Goals
2*Third party marks and brands are the property of their respective owners
AgendaAgenda
www.openib.orgwww.openib.org Linux Kernel InfrastructureLinux Kernel Infrastructure Linux Userspace InfrastructureLinux Userspace Infrastructure RDMA Enabled ApplicationsRDMA Enabled Applications InfiniBand and BeyondInfiniBand and Beyond
The InfiniBand Software SolutionThe InfiniBand Software Solution
3*Third party marks and brands are the property of their respective owners
OpenIB AllianceOpenIB Alliancewww.openib.org
•Improve interoperability
•Reduce the time and cost for deployment
•Improve interoperability
•Reduce the time and cost for deployment
Open-Source InfiniBand RDMA InfrastructureOpen-Source InfiniBand RDMA Infrastructure
4*Third party marks and brands are the property of their respective owners
OpenIB AllianceOpenIB Alliancewww.openib.org
…and the open source community!…and the open source community!
5*Third party marks and brands are the property of their respective owners
PhilosophyPhilosophy
Open source developmentOpen source development–Dual BSD / GPL LicenseDual BSD / GPL License
Integrated with OS infrastructureIntegrated with OS infrastructure Working with companion RDMA standardsWorking with companion RDMA standards AnyoneAnyone can contribute can contribute
www.openib.org
Code Quality,Code Quality, Stability,Stability, PerformancePerformance
6*Third party marks and brands are the property of their respective owners
AchievementsAchievementswww.openib.org
JunJun NovNov MarMar MayMay JulJul‘05‘04
Will support Windows & RDMA enabled Ethernet
Will support Windows & RDMA enabled Ethernet
Infiniband Leaders Form OpenIB Alliance
Infiniband Leaders Form OpenIB Alliance
Deliver a high-performance, scalable, and interoperable Infiniband software stack
Deliver a high-performance, scalable, and interoperable Infiniband software stack
“This will enable … a robust, resilient, and manageable infrastructure for data center, Grid, and HPC.”
“This will enable … a robust, resilient, and manageable infrastructure for data center, Grid, and HPC.”
“… accelerate the creation of a single Infiniband software stack”
“… accelerate the creation of a single Infiniband software stack” OpenIB Linux
release
OpenIB Linux release
7*Third party marks and brands are the property of their respective owners
Software ArchitectureSoftware Architecture
HCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
PMASMA SA ClientCM
HCA Driver
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
kDAPLIPoIB SRPSDP
iSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL RDMA App.
www.openib.org
HCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
PMASMA SA ClientCM
HCA DriverHCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
PMASMA SA ClientCM
HCA Driver
kDAPLIPoIB SRPSDP
iSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL RDMA App.
kDAPLIPoIB SRPSDP
iSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL RDMA App.
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
8*Third party marks and brands are the property of their respective owners
Linux Kernel InfrastructureLinux Kernel Infrastructure
HCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
kDAPL
PMASMA SA ClientCM
HCA Driver
IPoIB SRPSDPiSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
RDMA App.
9*Third party marks and brands are the property of their respective owners
OverviewOverview
Verbs API derived from VAPIVerbs API derived from VAPI Best of breed software solutionBest of breed software solution Support multiple HCASupport multiple HCA Optimized for Linux kernelOptimized for Linux kernel Support latest stable kernel releaseSupport latest stable kernel release
Linux Kernel Infrastructure
Complete InfiniBand infrastructureComplete InfiniBand infrastructure
10*Third party marks and brands are the property of their respective owners
ArchitectureArchitecture
HCA
HCA Driver
Linux Kernel Infrastructure
ib_mthca
Core InfiniBand Modules
ib_core Verbs
ib_mad
ib_sa
SA Client
ib_ipoibIPoIB
Linux 2.6.11
Linux 2.6.13CM
kDAPLSRPSDP
iSER NFS-R
Portals
Lustre
OpenIB Release
Development
External Project
MAD Services SMI GSI
PMASMAib_cm
RMPP
11*Third party marks and brands are the property of their respective owners
Linux Userspace InfrastructureLinux Userspace Infrastructure
HCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
kDAPL
PMASMA SA ClientCM
HCA Driver
IPoIB SRPSDPiSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
RDMA App.
12*Third party marks and brands are the property of their respective owners
OverviewOverview
Library exposes verbs to usermode clientsLibrary exposes verbs to usermode clients Secure userspace accessSecure userspace access Kernel bypass for speed path operationsKernel bypass for speed path operations RDMA to / from application memoryRDMA to / from application memory
Linux Userspace Infrastructure
Direct hardware access for applicationsDirect hardware access for applications
13*Third party marks and brands are the property of their respective owners
User-level InfiniBand Services
User-level Access Modules
user-space
kernel
ArchitectureArchitectureLinux Userspace Infrastructure
HCA
Core InfiniBand Modules
HCA Driver
User-level HCA Driver
Verbs
User Verbs
MAD Services
MAD Services
CM
CM
SA Client
SA Client
OpenSM OpenMPI uDAPL RDMA App.
openib.orgopenib.org
2.6.132.6.13
14*Third party marks and brands are the property of their respective ownersHCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
kDAPL
PMASMA SA ClientCM
HCA Driver
IPoIB SRPSDPiSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
RDMA App.
RDMA Enabled ApplicationsRDMA Enabled Applications
15*Third party marks and brands are the property of their respective owners
One Fabric, Many UsesOne Fabric, Many Uses
Low latency IPC for clusteringLow latency IPC for clustering Variety of high speed file and block Variety of high speed file and block
protocols, remote boot capable protocols, remote boot capable Bridging to external GbE and FCBridging to external GbE and FC
Multi-Service Switches
Management
Servers
ParallelFC / IB / IP
Storage
InfiniBand
IB HCA
IPC (MPI & DAPL)
Applications
Network (SDP & IPoIB)
Storage (File & Block)
InfiniBand Host Software Stack
RDMA Enabled Applications
One fabric has it allOne fabric has it all
16*Third party marks and brands are the property of their respective owners
Networking / IPC APIsNetworking / IPC APIsRDMA Enabled Applications
Tunnels IP packets over IB
Tunnels IP packets over IB
Application APIs unchanged
Application APIs unchanged
HCA
Core InfiniBand Modules
HCA Driver
Userspace HCA Driver
Userspace InfiniBand Services
user-space
kernel
IPC Application
NetPipe
Intel MPI
DAPL MPI
Sockets
IPoIB
IP
TCP
SDP
Perftest
OpenIB software enables applications
OpenIB software enables applications
Fabric agnostic RDMA API
Fabric agnostic RDMA API
Standard API for scientific clustering
Standard API for scientific clustering
Bypasses TCP/IP stack for direct
HW access
Bypasses TCP/IP stack for direct
HW access
17*Third party marks and brands are the property of their respective owners
Networking / IPC PerformanceNetworking / IPC PerformanceRDMA Enabled Applications
How well can apps perform?How well can apps perform?
Ciara VXB-7520J
3.4 GHz Dual EM64T Xeon3.4 GHz Dual EM64T Xeon 4 GB DDR4 GB DDR RedHat EL4.0RedHat EL4.0 OpenIB SVN 2833 backportedOpenIB SVN 2833 backported 10 Gb/sec (4X) - 8X PCI Express HCA10 Gb/sec (4X) - 8X PCI Express HCA
MTS2400
MTLP25208
18*Third party marks and brands are the property of their respective owners
Networking / IPC PerformanceNetworking / IPC Performance
Performance tests and ratings are measured using Performance tests and ratings are measured using specific computer systems and/or components and specific computer systems and/or components and reflect the approximate performance of Intel products reflect the approximate performance of Intel products as measured by those tests. Any difference in system as measured by those tests. Any difference in system hardware or software design or configuration may hardware or software design or configuration may affect actual performance. Buyers should consult affect actual performance. Buyers should consult other sources of information to evaluate the other sources of information to evaluate the performance of systems or components they are performance of systems or components they are considering purchasing.considering purchasing.
RDMA Enabled Applications
I.e. your results may varyI.e. your results may vary
19*Third party marks and brands are the property of their respective owners
Networking / IPC PerformanceNetworking / IPC Performance
0
5
10
15
20
25
30
usecs0
200
400
600
800
1000
MBps
RDMA Enabled Applications
Small message latencySmall message latency
IPoIBIPoIB SDPSDP MPIMPI
3.46 us3.46 us
924 MBps924 MBps
VerbsVerbs
Large message bandwidthLarge message bandwidth
20*Third party marks and brands are the property of their respective owners
InfiniBand Revolutionize InfiniBand Revolutionize StorageStorage
High-speed NAS and SAN on the same fabricHigh-speed NAS and SAN on the same fabric Scale out storage through clustering and Scale out storage through clustering and
virtualizationvirtualization
Server Racks A Bunch of low cost Storage Arrays
A lot of bandwidth
(InfiniBand)
Commodity Servers running NAS Gateways or Cluster File System
Switch
RDMA Enabled Applications
21*Third party marks and brands are the property of their respective owners
Storage in OpenIBStorage in OpenIBRDMA Enabled Applications
Storage
NICR-NICIB HCA
Switch
SCSI Midlayer
SRP
Core InfiniBand Services
HCA Driver NIC Driver
TCP/IP
iSCSI
RNIC Driver
kDAPL
iWarp Services
iSER
22*Third party marks and brands are the property of their respective owners
NFS/RDMANFS/RDMA
High-speed RDMA High-speed RDMA RPC transportRPC transport
Maps transparently to Maps transparently to NFS and applicationsNFS and applications
Reside over kDAPL Reside over kDAPL for fabric for fabric independenceindependence
NFS
RPC/XDR Switch
UDPTCPRDMA
RDMA Standard NIC
Direct Access
kDAPL
RDMA Enabled Applications
23*Third party marks and brands are the property of their respective owners
Xen Virtualization over IBXen Virtualization over IB
VM1
OS
VM2
OS
VM3
OS
GbE NIC FC HCA
Software based Network and Storage switching
XEN
StorageNetwork
Slower I/OSlower I/O No isolation No isolation Multiple cards and fabrics Multiple cards and fabrics
Direct HW access for I/ODirect HW access for I/O One 10Gb/s card for Network, One 10Gb/s card for Network,
Storage, and IPCStorage, and IPC Scale out using clustering (IPC)Scale out using clustering (IPC)
VM1
OS
VM2
OS
VM3
OS
IB HCA
XEN
IPC
Direct HW access
Xen with IBXen, the traditional way
RDMA Enabled Applications
24*Third party marks and brands are the property of their respective owners
Infiniband and BeyondInfiniband and Beyond
HCA
Core InfiniBand Modules
Verbs
MAD Services
SMI GSI RMPP
kDAPL
PMASMA SA ClientCM
HCA Driver
IPoIB SRPSDPiSER NFS-R
Portals
Lustre
OpenSM OpenMPI uDAPL
User-level HCA Driver
User-level InfiniBand Services
VerbsMAD Services SA ClientCM
User-level Access Modules
CM SA ClientMAD Services
user-space
kernel
User Verbs
RDMA App.
What’s left?What’s left?
25*Third party marks and brands are the property of their respective owners
Future EffortsFuture Efforts
Scale out testingScale out testing Scalability enhancementsScalability enhancements Integrate upper-layer protocol softwareIntegrate upper-layer protocol software Windows software releaseWindows software release Xen supportXen support
InfiniBand and Beyond
26*Third party marks and brands are the property of their respective owners
Companion TechnologiesCompanion Technologies
Working with OpenRDMA GroupWorking with OpenRDMA Group Common RDMA interface for LinuxCommon RDMA interface for Linux Support InfiniBand and iWarpSupport InfiniBand and iWarp
InfiniBand and Beyond
HCA
HCA Driver
RDMA Application
RNIC
InfiniBand Services
RNIC Driver
iWarp Services
RDMA Verbs
RDMA Application Library
Connection Mgmt
InfiniBand Verbs
HCA
HCA Driver
InfiniBand Services
SA ClientCM
RDMA Application
kDAPL
HCA
HCA Driver
RDMA Application
InfiniBand Services
RDMA Verbs
RDMA Application Library
HCA
HCA Driver
RDMA Application
InfiniBand Services
RDMA Verbs
RDMA Application Library
Connection Mgmt
27*Third party marks and brands are the property of their respective owners
SummarySummary
Solid InfiniBand infrastructureSolid InfiniBand infrastructure Available in Linux today, Windows soonAvailable in Linux today, Windows soon Low latency, high bandwidth applicationsLow latency, high bandwidth applications
OpenIB – Open-Source Infiniband RDMA Infrastructure
OpenIB = InfiniBand softwareOpenIB = InfiniBand software
28*Third party marks and brands are the property of their respective owners
Call to ActionCall to Action
Join the open source effortJoin the open source effort Request OpenIB software from your Linux Request OpenIB software from your Linux
distributordistributor Ask your IB vendor about OpenIB supportAsk your IB vendor about OpenIB support
InfiniBand and Beyond
Join OpenIBJoin OpenIB
29*Third party marks and brands are the property of their respective owners
Please fill out the Please fill out the Session Evaluation Session Evaluation Form.Form.
Thank You!Thank You!
30*Third party marks and brands are the property of their respective owners
GlossaryGlossary CM – communication managerCM – communication manager DAPL – direct access provider libraryDAPL – direct access provider library GSI – general services agentGSI – general services agent HCA – host channel adapterHCA – host channel adapter IPC – interprocess communicationIPC – interprocess communication IPoIB – IP over InfinibandIPoIB – IP over Infiniband iSER – iSCSI extensions for RDMAiSER – iSCSI extensions for RDMA PMA – performance management agentPMA – performance management agent SDP – sockets direct protocolSDP – sockets direct protocol SM – subnet managerSM – subnet manager SMI – subnet management interfaceSMI – subnet management interface SRP – SCSI RDMA procotolSRP – SCSI RDMA procotol
OpenIB – Open-Source Infiniband RDMA Infrastructure
31*Third party marks and brands are the property of their respective owners
Additional Resources for this Additional Resources for this SessionSession Session presentation can be downloaded Session presentation can be downloaded
from the IDF web site – when prompted from the IDF web site – when prompted enter:enter:
–Username: idfUsername: idf
–Password: fall2005Password: fall2005
More web based info: www.openib.orgMore web based info: www.openib.org