uc santa cruz impact of failure on interconnection networks for large storage systems qin xin, ethan...

UC Santa Cruz

Impact of Failure on Interconnection Networks for Large Storage Systems

Qin Xin, Ethan L. Miller, Thomas J. E. Schwarz*, Darrell D. E. Long

Storage Systems Research Center University of California, Santa Cruz

*Department of Computer EngineeringSanta Clara University

MSST'05 24/13/05

Problems

Reliability is a real concern for large storage systems• A large number of components • Complex interconnections• Human errors

Robust network interconnection is desired• Failures of switches, routers, and network links are

common• Organization of nodes and links is crucial

Scale makes things complicated: petabyte storage system --10,000s disks, 100s routers,100,000s links

• Robust network topology Tolerant multi-points of failures without noticeable

performance degradation

MSST'05 34/13/05

Contributions

Examine various network topologies• Butterfly networks• Mesh structure• Hypercube• Tradeoffs among them: cost, robustness,

performance. Estimate the impacts of link and node failures

by simulation• I/O path connectivity• Number of hops in varied degraded modes• Failure impact to its neighborhood

MSST'05 44/13/05

Network Topology for Storage Systems

Butterfly networks• Reasonable cost• Poor robustness

Mesh• Brick-structured, good

robustness• Relatively long path

Hypercube• Superior robustness• Complex, costly

2

3

4

server 1

server 2

server 3

server 4

router

router

router

router

switch

switch

switch

switch

switch

switch

switch

switch

server

router

server

0 1

3 4

2

5

6 7 8

router 1

router 2

MSST'05 54/13/05

A case study: Failure Scenarios

server

0 1

3 4

2

5

6 7 8

router 1

router 2

server

0 1

3 4

2

5

6 7 8

router 1

router 2

server

0 1

3 4

2

5

6 7 8

router 1

router 2

server

0 1

3 4

2

5

6 7 8

router 1

router 2

MSST'05 64/13/05

Simulated Results: Connectivity

I/O path connectivity• Measured in “nines”: -log10(1-P), P: m/n, m: # of good

requests, n: # of total requests. (3 nines = 99.9%)• Poor connectively when the connection nodes are lost. • Hypercube and mesh are more robust in connectivity than

butterfly networks.

MSST'05 74/13/05

Simulation Results: Neighborhoodmesh

hypercube Failure impact on network

neighborhood• Neighbors suffer in presence of

failures.• Hypercube outperforms butterfly &

mesh structure in avg. I/O load.

butterfly

MSST'05 84/13/05

Summary

A well-chosen topology can provide robust interconnection.

Neighbors around failures suffer much more than average.

Hypercube structure provides better interconnection in degraded modes than butterfly and mesh structure.

MSST'05 94/13/05

Questions?

For further information, please visit:• http://ssrc.cse.ucsc.edu/• http://www.cs.ucsc.edu/~qxin/

Thanks to SSRC faculty, students, affiliates, and alumni Thanks to our sponsors

• Lawrence Livermore National Laboratory, Los Alamos National Laboratory, Sandia National Laboratory

• Engenio Information Technologies, HP Lab, Hitachi Global Storage Technologies, IBM Research, Intel, LSI Logic, Microsoft Research, Network Appliance, and Veritas

• NSF, USENIX

MSST'05 104/13/05

Big Picture: A Petabyte-Scale Storage System

High Speed NetworkWAN

...

... ...

...

MetadataServer

...

MetadataServer

SuperComputer

SuperComputer

devicedevice

Client

Client

10s10,000s

1000s

100,

000s

Potential for data loss• Thousands of disks• Petabytes of data