1j. kim web science & technology forum enabling hardware technology for web science john kim...

14
1 J. Kim Web Science & Technology Forum Enabling Hardware Technology for Web Science John Kim Department of Computer Science KAIST

Upload: walter-wilson

Post on 27-Dec-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

1J. Kim Web Science & Technology Forum

Enabling Hardware Technology for Web Science

John Kim

Department of Computer Science

KAIST

2J. Kim

Computing System

Web Science & Technology Forum

3J. Kim

Moore’s Law

Web Science & Technology Forum

4J. Kim

Increasing number of cores

Rethinking the Design of Interconnection Networks

1975 1980 1985 1990 1995 2000 2005 2010 20150.512

5.12

51.2

512

Year

# Co

res

8086 80286 80386 80486 PentiumPentium Pro

P-II P-III

P-4Itanium

Power4

MIT-RAW

UltraSPARC IVPentium-D

Xeon Brisbane

Niagara

Core 2 Quad

Turion X2Core 2 DuoDenmark

Teraflops

Larrabee

TILE64

Sun Rock

Cell

XeonAgena Barcelona

Toliman

Core i7

IstanbulBecktonMagny-Cours

5J. Kim

Challenges for Future Hardware

• Energy-Efficient Mobile Systems– Need to extend battery life– Provide the performance capability of a laptop with a smaller

power budget– Exploit parallelisms available in future multi/many-core

processors

• Scalable Datacenter– Future datacenters will continue to increase in the number of

servers interconnected– Need to minimize energy consumption

Web Science & Technology Forum

6J. Kim

Energy Consumption

• Datacenters have a huge electricity bill– e.g.) US datacenters will consume 100 billion kWh at a cost

of $7.4 billion per year [EPA report]

• Total cost of ownership is no longer dominated by the hardware but by

Energy-Efficient Datacenters

Source: C. Belady “In the data center, power and cooling costs more than the it equipment it supports”

7J. Kim

Which costs more?

Web Science & Technology Forum

20mm

8J. Kim

Supercomputers to Datacenters

Web Science & Technology Forum

9J. Kim

Can we leverage the network used in high-performance computing for

datacenter networks?

funded in part by Microsoft Research Asia

Web Science & Technology Forum

10J. Kim

Existing approaches

• Tree– Oversubscribed– Bottlenecked root

• Fat tree– 1:1 subscription– Costly as large number

of switches

With the number of servers growing

Container0 Container1 ...

TOR

switches

TOR

switches

With the number of servers growing

Scale up

Scale out

High end router

10slide from C. Guo

11J. KimRethinking the Design of Interconnection

Networks

Flattened Butterfly

John Kim, William Dally, Dennis Abts, “Flattened Butterfly: Cost-Efficient High-Radix Networks” ISCA 2007

12J. Kim

Dragonfly Topology [ISCA’08]

• Increase scalability by using a collection of routers as a ”supernode”

• Leverage the packaging hierarchy found in systems to match the hierarchical network of the topology?

• What is the challenge of mapping this topology?

Adaptive Routing

Web Science & Technology Forum

13J. Kim

Packaging Hierarchy

Web Science & Technology Forum

Rackable Systems Container

14J. Kim

Load-Balanced Routing

• Load-balancing required in both inter- and intra-container routing

• Randomized routing can be leveraged to load balance the channels

• Adaptive routing need to determine whether to route using minimal routing and nonminimal routing

Approach :

- Leverage the centralized management capability in datacenters and provide a centralized adaptive routing.

- Focus on load-balancing the “elephant” flows while other short-lived flows can be routed minimally

Web Science & Technology Forum