High Availability in Clustered Multimedia Servers

1
High Availability in Clustered Multimedia Servers

• Renu Tewari Daniel M. Dias Rajat Mukherjee
  Harrick M. Vin

2
Topics

• Problem high availability in clustered
  multimedia servers
• Schemes for high availability
• Details of some schemes for high availability
• Simulation
• Cost performance analysis
• Conclusion

3
The problem the author addressed

• High availability in clustered multimedia servers
• Clustered multimedia servers consists of a set of
  processing node, each with a local disk array,
  connected by a high bandwidth switch or network.
  High availability requires the servers can
  provide continuous delivery in the presence of
  failures, including disk failure and node failure.

4
Architecture of clustered multimedia servers
Front end performs delivery. Back end provides
storage.
5
Architecture of clustered multimedia servers

• Front end fails
• Stream can be resumed from another delivery node.
• Back end fails
• Has system-wide effect. It is the major concern
  in this paper.

6
Schemes for high availability

• Mirroring
• Disk level mirroring
• Block level mirroring

7
Schemes for high availability

• Twin-tailing/Multi Tailing
• Used to handle node failure. A buddy node can
  access the others disks in case of failure.

8
Schemes for high availability

• Software raid
• Sequential parity placement
• Random parity placement

9
Sequential parity placement

• Disk numbered i, is attached to a node numbered i
  mod N, Nnumber of nodes
• Parity group, if size5, DDDDPDDDDP

10
Sequential parity placement
11
Sequential parity placement
12
Random parity placement

• Constraint
• Blocks belonging to the same parity group are not
  placed on the same disk or on any disk on the
  same node.

13
Random parity placement
14
Some simulation results
15
Some simulation results
16
Comparison between spp and rpp

• According to the simulation, RPPs performance is
  better than SPP.
• RPP needs much more meta-data.
• Increasing the read-ahead buffer size,with RPP,
  the loss can be reduced substantially.
• To decrease the amount of meta data of RPP, SPP
  with multiple strides can be used.

17
Cost analytical model

• Queue theory
• Each disk behaves like an M/D/1 queue.

18
Conclusions

• Sequential placement of parity can only balance
  the space and bandwidth utilization of all disks
  during normal operation.
• Balanced random placement of parity can achieve
  during failure and normal operation.
• Mirroring is cost effective only when missing the
  real time constraints during failure is more
  expensive than the extra disk capacity.
• Larger parity group sizes have smaller disk space
  overhead but larger memory costs for buffering.
• Tighter loss criteria require small parity group
  size.

19
Questions

• Name 3 schemes for high availability.
• In sequential parity placement, why should you
  choose the parity group size to be relatively
  prime to the number of disks?