View
2
Download
0
Category
Preview:
Citation preview
Greg PapadopoulosEVP and CTOSun Microsystems, Inc.
Innovation Matters
Takeaways● Business demands fuel IT demand
– Still in the first phases of network computing
● Decision point– Mask complexity through people or eliminate
through engineering
● Innovation matters
The Revolution
10M
1B
10B
1T
100B
1990 2000 2010 2020
Computers People Everything
Cost toConnect
ThingsConnected
$100
$1
1¢
RFID
The Revolution
10M
1B
10B
1T
100B
1990 2010 2020
Computers People Everything
Connectivity
ConnectivityCommunity
Community
Awar
enes
s
Awar
enes
s
2000
Bottom Line● Still in the first phases of networking
– Expect the number of networked things to grow by another factor of 1M
– New value from automating still-manual processes
● Expect radical transformation of SW and systems– Systems, in particular, are at a watershed– Mask complexity through people or eliminate
through engineering
ProcessorProcessor
O/SO/S
App
Shrink-Wrap
SolarisSparc
Linuxx86
VMSVAX
MVS370
AIXPower
MAC OSPower
IrixMIPS
HP-UXPrecision
DGUX88K
OS400AS400
Windowsx86
UltrixAlphaAegis
68K
The New Software
Application scales from 1 to perhaps 100's of users
For Desktop:1 copy SW = 1 copy HW
Shrink-Wrap
SolarisSparc
Linuxx86
VMSVAX
MVS370
AIXPower
MAC OSPower
HP-UXPrecision
DGUX88K
OS400AS400
Comp
MWMW
Comp
MWMW
Comp
MWMW
Comp
MWMW
Windowsx86
UltrixAlphaAegis
68K
NetworkedComponents
The New Software
ProcessorProcessor
O/SO/S
App
Service must scaleto 10,000's or millionsof simultaneous users!
IrixMIPS
TheThe Java JavaTMTM Platform Platform
The WindowsThe WindowsPlatformPlatform
SolarisSparc
Linuxx86
VMSVAX
MVS370
AIXPower
MAC OSPower
IrixMIPS
HP-UXPrecision
DGUX88K
OS400AS400
Windowsx86
UltrixAlphaAegis
68K
XML
The New Software
ProcessorProcessor
O/SO/S
App Comp
MWMW
Comp
MWMW
Comp
MWMW
Comp
MWMW
The Operations Challenge
Was...● Availability● Performance● Scale
Price/Price/PerformancePerformance
The Operations Challenge
Becoming...● Service Level● Efficiency● Security
Total Cost ofTotal Cost ofService LevelService Level
Price/Price/PerformancePerformance
Was...● Availability● Performance● Scale
The Emperor Has No Clothes● Network computing is wildly complex:
no discernible economy of scale● 75+% of IT budgets going to people +
operations● 15–20% typical utilization● Out of the $2.4T WW ICT spending only
a tiny fraction goes to HW/SW systems
Economy of Scale
Bad
Scale
Cost
Economy of Scale
Better
Bad
Scale
Cost
Mask ComplexityMask Complexityoror
Engineer It AwayEngineer It Away
Decision Point
Old Systems Are Components in the New One
Server connectedto networks
Old Systems Are Components in the New One
Server connectedto networks
Datacenter systembuilt from networks
What Is O/S for This New System?
NOT these!
What Is O/S for This New System?
Grid Computing Today at Sun13,500 CPUs in 3 cities for processor design
22 years compute time/day98% average CPU usage 24/7/365
● SPARC microprocessor design facilities in Sunnyvale, Austin, Burlington
● Sites use a departmental Cluster Grid approach normally
● Workload is distributed across all sites in an Enterprise Grid approach to meet peak load requirements
22
THROUGHPUT COMPUTING ● New design approach for the UltraSPARC®
processor family● Eclipses throughput of today's processors by
a magnitude of up to 15–30x● Quantum reduction in the cost of network
computing—without software disruption
Network Computing Is Thread RichWeb services, JavaTM applications, database transactions, ERP . . .
Moore’s LawA fraction of the die can already build a good processor core; how am I going to use a billion transistors?
Worsening Memory LatencyIt’s approaching 1000sof CPU cycles! Friend or foe?
Forcing a rethinking of processor architecture – modularity, less is more, time-to-market
Growing Complexityof Processor Design
The Big Bang Is Happening—Four Converging Trends
Memory BottleneckRelative Performance
10000
1
1990 1995 2005 1980
1000
100
10
1985 2000
2x Every 2 YearsCPU Frequency
Memory BottleneckRelative Performance
10000
1
1990 1995 2005 1980
1000
100
10
1985 2000
2x Every 6 Years
2x Every 2 Years
Gap
CPU FrequencyDRAM Speeds
C MC MC M
Memory Latency Compute
Time Memory Latency Compute
HURRYHURRYUP ANDUP ANDWAIT!WAIT!
Note: Up to 75% Cycles Waiting for MemoryNote: Up to 75% Cycles Waiting for Memory
Typical Complex HighFrequency Processor
Thread
Memory Latency Compute
Time
Time Saved
Memory Latency Compute
C M C M C MThread
*Source: Microprocessor Report, Richard Sites, It's the Memory, Stupid!, August 1996
Chip Multithreading (CMT)
Memory Latency Compute
Time Memory Latency Compute
C MC MC MThread 1
Thread 2 C MC MC M
Thread 3 C MC MC M
Thread 4 C MC MC M
Core 1
Memory Latency Compute
CMT—Multiple Multithreaded Cores
Thread 4Thread 3Thread 2Thread 1
Core 2Thread 4Thread 3Thread 2Thread 1
Core 3Thread 4Thread 3Thread 2Thread 1
Thread 4Thread 3Thread 2Thread 1
Core 4
Core 5Thread 4Thread 3Thread 2Thread 1
Core 6Thread 4Thread 3Thread 2Thread 1
Core 7Thread 4Thread 3Thread 2Thread 1
Core 8Thread 4Thread 3Thread 2Thread 1
Time
Gen. 1 CMT SingleThreadedProcessor
Gen. 2 CMT15X
3X3X
Gen. 1 CMT SingleThreadedProcessor
3X3X2X
Gen. 3 CMT30X
4X5X
Gen. 2 CMT15X
SingleThreadedProcessor
3X
RelativePerformance
30
25
20
15
Throughput Computing
10
5
Time Network FacingData Facing
0 Today
UltraSPARC® IV1st GENERATION CMT PROCESSOR FOR MID/HIGH END SYSTEMS
● First milestone in Sun's Throughput Computing roadmap
● Up to twice the throughput of today's UltraSPARC III processor
● Protects customer investment in hardware, software and staff– Upgrade path– Binary compatibility across
time and product line
The Three Waves of SPARC Innovation
Pric
e/Pe
rfor
man
ce
1980 20001990 2010
RISC
SMPCMT
Takeaways● Business demands fuel IT demand
– Exponential demand as everything becomes connected
● Decision point– Eliminate complexity through engineering
● Innovation from Sun– In Software– In Systems
Greg PapadopoulosEVP and CTOSun Microsystems, Inc.
Innovation Matters
Recommended