mellanox technologies maximize cluster performance and...
TRANSCRIPT
Gilad Shainer, [email protected], 2007
Mellanox TechnologiesMaximize Cluster Performance and Productivity
2 Mellanox Technologies
Enterprise Data CentersEnterprise Data Centers
EmbeddedEmbedded
Mellanox TechnologiesApplicationsHardware OEMs End-Users
HighHigh--Performance ComputingPerformance Computing
InfiniBand and Ethernet
ServersAnd Blades
Embedded
Switches
Storage
Interconnect: A Competitive Advantage
3 Mellanox Technologies
Connecting The Most Powerful Clusters
4500 server nodes
1300 servernodes
1280 server nodes
2300 server nodes
1400 server nodes
96 server nodes
4 Mellanox Technologies
0
100
200
300
400
500
8 Cores 16 Cores 32 Cores
Time to Compute 10 Time Steps
But Not Only The Most Powerful ClustersSikorsky CH-53K program• 24 nodes, dual core CPUs• Reducing simulations
duration from 4 days to several hours
Personal supercomputing (4-5 nodes)• Wolfram Air Pollution Simulation • Maximum utilization and efficiency
Seconds
HP
TYAN
5 Mellanox Technologies
Clusters• Commodity• Very flexible
InfiniBand Clusters• Maximum performance• Scalability, large-scale clusters• Multiple I/O traffics• Flexible and easy to manage
Proprietary Systems• Expensive• Not flexible
MellanoxInfiniBand
Off-the-ShelfStorage
CommodityServers
InfiniBand For Clustering
IndustryMegatrend
6 Mellanox Technologies
InfiniBand For I/O Growth Demands
Multi-core CPUs mandating 10Gb/s+ connectivity
More applications per More applications per server and I/Oserver and I/O
More traffic per I/O More traffic per I/O with server I/O with server I/O consolidationconsolidation
I/O capacity per server I/O capacity per server dictated by the most dictated by the most
demanding appsdemanding apps
10Gb/s+ 10Gb/s+ connectivity connectivity
for each for each serverserver
Multi Core CPUs
SAN Adoption
SharedResources
More applicationsMore applicationsdemand realdemand real--time time
responseresponse
LatencyRace
More applications per More applications per server and I/Oserver and I/O
More traffic per I/O More traffic per I/O with server I/O with server I/O consolidationconsolidation
I/O capacity per server I/O capacity per server dictated by the most dictated by the most
demanding appsdemanding apps
10Gb/s+ 10Gb/s+ connectivity connectivity
for each for each serverserver
Multi Core CPUs
SAN Adoption
SharedResources
More applicationsMore applicationsdemand realdemand real--time time
responseresponse
LatencyRace
7 Mellanox Technologies
InfiniBand For Virtualization
Hardware-based I/O virtualization • Multiple VMs, multiple traffic types per VM
Supports current and future servers• AMD and Intel IOV, PCI-SIG IOV
Better resource utilization• Frees up CPU through hypervisor offload• Enable significantly more VMs per CPU
Native OS performance• VMs enjoy native InfiniBand performance
ConnectX InfiniBand HCA
VirtualMachine 1
HCA Driver
…
DMA Remapping
Memory
VirtualMachine 1
HCA Driver
VirtualMachine 1
HCA Driver
Physical Server
Virtual Machine Monitor
VMM OffloadFunctions
8 Mellanox Technologies
TOP500 Interconnect Trends
Explosive growth of InfiniBand 20Gb/s• 55 clusters = 42% of the InfiniBand clusters
Top500 Interconnect Penetration
0
20
40
60
80
100
120
140
Year 1
Year 2
Year 3
Year 4
Year 5
Year 6
Year 7
Year 8
Year of Introduction
# of
Clu
ster
s
InfiniBand Ethernet
Ethernet year 1 – June 1996InfiniBand year 1 – June 2003
InfiniBand adoption is faster then Ethernet
InfiniBand In The Top500
0
20
40
60
80
100
120
140
June 05 Nov 05 June 06 Nov 06 June 07
Num
ber o
f Clu
ster
s
IB SDR IB DDR
LANL
SANDIA
AERONAUTICAL SYSTEMS CENTER
Top500 Interconnect Trends
020406080
100120140160180200220240260
Num
ber o
f Clu
ster
s
Jun-05 Nov-05 Jun-06 Nov-06 Jun-07InfiniBand Myrinet GigE
132
47
207
Top500 Interconnect Trends
020406080
100120140160180200220240260
Num
ber o
f Clu
ster
s
Jun-05 Nov-05 Jun-06 Nov-06 Jun-07InfiniBand Myrinet GigE
132
47
207
Growth Rate:
InfiniBand: +230%GigE: -19%
9 Mellanox Technologies
InfiniBand Value Proposition
Automotive Oil and Gas Fluid Dynamics
* Partial List
3X simulation efficiency increaseCar crash simulations eliminates the need for physical testing
Reduce reservoir modeling simulation runtime up to 55%Improves interpretation and modeling accuracy
3X performance improvement and near linear scalingIntensive simulation for computer-aided engineering
5X the bandwidth for real-time color gradingHigh-resolution commercials and feature films
4X improvement for photomaskmanufacturing - CATS
Monte Carlo simulation, astronomy, bioinformatics, chemistry and drug researchAccelerate parallel execution of matrix operations
Oil and GasAutomotive Fluid Dynamics
Electronic Design AutomationDigital Media Computational Science
10 Mellanox Technologies
Mellanox Product Roadmap
2000 2003 2004 2005 2006 20072001 2002
40Gb/s Total
160Gb/s Total
Two 10Gb/s
480, 960Gb/s Total
Two 10/20Gb/s IB
One 10/20Gb/s
1st Gen.
2nd Generation
3rd Generation
4th Generation
Adapter
Adapter
Adapter
Adapter + Switch
Switch
Switch
*Adapter Card Products Based on Adapter Silicon Not Shown
2.0 AdapterTwo 10/20/40 Gb/s InfiniBand or Two 1/10Gb/s Ethernet
11 Mellanox Technologies
Leading InfiniBand and 10GigE Adapters
Single-chip server and storage adapters• Optimized cost, power, footprint, reliability
Highest performing InfiniBand adapters• 20Gb/s (40Gb/s in 2008), 1us application latency
Highest performing 10GigE NICs• 17.6Gb/s throughput, < 7us application latency• Supports OpenFabrics RDMA software stacks
Multi-core CPU optimizedVirtualization accelerationCombination IB/Ethernet 4Q07First adapters to support PCI Express 2.0 10 & 20Gb/s InfiniBand
10 Gigabit Ethernet(Copper)
ConnectX IB InfiniBand HCA
ConnectX EN Ethernet MAC/NIC
ConnectX MP Multi-Protocol Adapter
12 Mellanox Technologies
ConnectX Multi Core PerformanceConnectX MPI Latency - Multi-core Scaling
0
2
4
6
1 2 3 4 5 6 7 8
# of CPU cores
Late
ncy
(use
c)
*Optimizations on going
IPoIB-CM ConnectX IB - SDR, DDR PCIe Gen1, DDR PCIe Gen2
0
500
1000
1500
20002 4 8 16 32 64 128
256
512
1024
2048
4096
8192
1638
432
768
6553
613
1072
2621
4452
4288
1048
576
Bytes
Ban
dwid
th (M
B/s
)
IB SDR PCIe Gen1 IB DDR PCIe Gen1 IB DDR PCIe Gen2
Sockets (TCP)
MPI (message passing)
13 Mellanox Technologies
InfiniBand scalability - car2car
0
10
1 2 4 8 16
Number of nodes
Perf
orm
ance
Superior Application Productivity
PAM-CRASH
05000
10000150002000025000300003500040000
16 CPUs 32 CPUs 64 CPUs
Elap
sed
Tim
e (s
ec)
GigE InfiniBand
FLUENT 6.3Beta - FL5L3 case
0
500
1000
1500
2000
2500
0 20 40 60 80 100 120 140
Number of Cores
Flue
nt P
erfo
man
ce R
atin
g
Mellanox InfiniBand Gigabit Ethernet
HP C-Class Blade System with Mellanox 20Gb/s InfiniBand I/O
14 Mellanox Technologies
Superior Application Performance
Mellanox InfiniBand versus QLogic InfiniPath - neon_refined_revised
50%
60%
70%
80%
90%
100%
2 nodes 4 nodes 8 nodes
Nor
mal
ized
Run
time
QLogic InfiniPath Mellanox InfiniBandLower isbetter
100
120
140
160
180
200
Tim
e (se
c)
Vector Distribution Benchmark
GigE InfiniBandLower is better
57%
0
100
200
300
400
500
8 Cores 16 Cores 32 Cores
Time to Compute 10 Time Steps
Systems: Intel Woodcrest 3.0GHz Interconnects: Mellanox InfiniBand SDR, QLogic InfiniPath
Mellanox InfiniBand – full CPU offload• Transport offload• RDMA• Robustness• Scalability• Efficiency
CPU runs application, not network
15 Mellanox Technologies
Mellanox Cluster Center
http://www.mellanox.com/applications/clustercenter.php
Neptune cluster • 32 nodes • Dual core AMD Opteron CPUs
Helios cluster • 32 nodes• Quad core Intel Clovertown CPUs
Vulcan cluster – coming soon• 32 nodes• Quad core AMD Barcelona CPUs
Utilizing “Fat Tree” network architecture (CBB)• Non-blocking switch topology • Non-blocking bandwidth
ConnectX InfiniBand 20Gb/sInfiniBand based storage
• InfiniBand storage• NFS over RDMA, SRP
16 Mellanox Technologies
SummaryMarket-wide adoption of InfiniBand • Servers/blades, storage and switch systems• Data Centers, High-Performance Computing, Embedded• Performance, Price, Power, Reliable, Efficient, Scalable
4th Generation adapter - connectivity to InfiniBand and Ethernet• Market leading performance, capabilities and flexibility
Driving key trends in the market• Clustering/blades, low-latency, I/O consolidation, multi-core, virtualization
Fibre Channel Storage
iSCSI StorageConverged Architecture with ConnectX – I/O Consolidation
InfiniBand Storage
InfiniBand orEthernet
Low Cost Bridge
InfiniBand orEthernet
Thank You
17