38
Distributed Cluster Repair for OceanStore Irena Nadjakova and Arindam Chakrabarti Acknowledgements: Hakim Weatherspoon John Kubiatowicz

Distributed Cluster Repair for OceanStore

  • Upload
    trevet

  • View
    32

  • Download
    0

Embed Size (px)

DESCRIPTION

Distributed Cluster Repair for OceanStore. Irena Nadjakova and Arindam Chakrabarti Acknowledgements: Hakim Weatherspoon John Kubiatowicz. OceanStore Overview. Data Storage Utility Robustness Security Durability High availability Global-scale. Where our project fits in. Durability - PowerPoint PPT Presentation

Citation preview

Page 1: Distributed Cluster Repair for OceanStore

Distributed Cluster Repair for OceanStore

Irena Nadjakova and Arindam Chakrabarti

Acknowledgements:Hakim Weatherspoon

John Kubiatowicz

Page 2: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 2

OceanStore Overview

Data Storage Utility• Robustness• Security• Durability • High availability • Global-scale

Page 3: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 3

Where our project fits in

Durability• Automatic version-management• Highly redundant erasure-coding• Massive dissemination of

fragments on machines with highly uncorrelated availability.

Page 4: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 4

The internet

Page 5: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 5

Choosing locations for storing a fragment

Page 6: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 6

Choosing locations for storing a fragment

Page 7: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 7

Choosing locations for storing a fragment

Page 8: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 8

Clustering

Page 9: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 9

Clustering

Page 10: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 10

OceanStore solution

• Availability of each machine tracked over time.

• Machines that have very little availability are not used for fragment storage.

• Distance between each pair of machines computed. (high mutual information ) close)

• Cluster the machines into chunks based on this distance using normalized cuts.

• All the computation is done on one central computer (Cluster Server).

Page 11: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 11

OceanStore solution

• Machines that are highly correlated in availability are in same cluster.

• Machines in separate clusters have low correlation in availability.

• When a node needs to store replica fragments, it requests cluster information from the cluster server and uses it to send each fragment to k nodes: one from each of k different clusters.

Page 12: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 12

Cluster creation

• Needs centralized computation.• Can we do it in a distributed manner ?• NCuts is one stumbling block. It seems to

need the entire graph.• Having to pull the cluster info from one

central cluster server: single point of failure• Can we have a “Distributed NCuts” algo to

look at subgraphs ? How to make subgraphs? Do we need to know the entire graph to decide how to divide it into pieces ?

Page 13: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 13

Distributed clustering

Page 14: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 14

Distributed clustering

Page 15: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 15

Distributed clustering

Page 16: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 16

Distributed clustering

Page 17: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 17

Distributed clustering

Page 18: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 18

Initial idea

• We run the centralized algorithm once for some time period (chose 73 days) to generate some initial clustering (expensive!)

• We distribute the machines among some f cluster servers– Each has a smaller subset of size num of the

initial machines– Keeping the initial clustering proportions for

each node– Each machine occurs in approximately

equal number of cluster servers

Page 19: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 19

Initial idea (cont)

• Now we can afford to recluster the machines on each server frequently to keep up with the network changes.– Chose to do it once every 30 days for

the simulation purposes, but can easily be done a lot more often

Page 20: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 20

Evaluation

• To see how well this does, we want to compare it with the original global algorithm, run in the same time period.

• Metric – the average mutual informationI(x,y) = P(x,y) log P(x,y)/P(x)P(y)

– Average MI for a single server is just the average of the mutual information between pairs of machines in different clusters on the server

– On multiple servers, we compute the above on every server, then average among servers

Page 21: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 21

Simulating Network Evolution Dynamics

• We have availability data for 1000 machines for a period of 73 days.

• We use it to simulate the behavior of a network with 1000 machines over a period of 730 days = 2 years.

• We simulate networks with varying evolution characteristics to evaluate the robustness of our distributed cluster repair algorithm.

Page 22: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 22

Simulating Network Evolution Dynamics

Qualities of a good network:• Maybe server availability (AV) should

not vary drastically in the future ?• Maybe average server repair time

(MTTR) should not vary drastically ?• Maybe mean time to failure (MTTF)

should not vary drastically ?• Maybe failure correlations (FCOR)

should also not vary drastically ?

Page 23: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 23

NS Algo 1: Sanity Check 1

Global déjà vu• Maintains AV, MTTF, MTTR, FCOR• Simulates a well-behaved network.• Our distributed update algorithm

should do very well on this.

Page 24: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 24

NS Algo 2: Acid Test 1

Local déjà vu• Maintains AV, MTTF, MTTR, but not

FCOR.

Page 25: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 25

NS Algo 3: Acid Test 2

Births and Deaths• Maintains AV, MTTF, MTTR, and FCOR,

but only for some nodes, and for some time.

• Nodes are taken off (die) the network or are added to (born) the network at certain times. When they are actually on the network, they maintain their AV, MTTF, MTTR, FCOR.

Page 26: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 26

NS Algo 4: Acid Test 3

Noisy Global déjà vu• Maintains AV, MTTF, MTTR, FCOR

to a large extent, but adds some Gaussian noise, representing the variations that may be observed in a real network.

Page 27: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 27

NS Algo 5: Acid Test 4

Noisy Local déjà vu• Maintains AV, MTTF, MTTR, but not

FCOR, and also adds some Gaussian noise representing the variations that may be observed in a real network.

• Does our algorithm do well in this situation ? If yes, how robust is it to noise ?

Page 28: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 28

Page 29: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 29

Page 30: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 30

Page 31: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 31

Page 32: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 32

Still problems

• Initial clustering is expensive• What happens if we don’t use it?

Page 33: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 33

Page 34: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 34

How to fix this?

• Randomly distribute machines to the servers

• Perform local clustering• Find the ‘unwanted’ elements (highest

mutual information with the rest on this node)

• Exchange them with ‘unwanted’ elements of another cluster to which the first ones are least correlated

• Communication overhead is low; most computation can proceed without a lot of communication

Page 35: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 35

Under development…

• What we have so far is a scheme that – picks a server at random– finds a few unwanted elements– exchanges those with the same number of

unwanted elements of another server – picked at random, or having the best correlation with the unwanted elements of the first server

• The percentage improvement is small so far – 0.4%-1.5% for the first 5 or so runs. It falls off afterwards.

Page 36: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 36

How to improve

• Exchange more than 1 machines ?• Run for several generations ?• It may be even better to just

randomly exchange machines, as long as the overall average mutual information of the distributed cluster decreases.

Page 37: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 37

Summary of achievements

• Towards getting rid of expensive centralized cluster creation

• Scalable distributed cluster management scheme

Page 38: Distributed Cluster Repair for OceanStore

Dec 9, 2003 Distributed Cluster Repair for OceanStore 38

Thanks for listening !

Acknowledgements:

Hakim WeatherspoonJohn Kubiatowicz