Information Technology Reference
In-Depth Information
BS (PRT)
BS (FEC)
2
1.5
1
SSS (PRT)
BS (non-FT)
SSS ( PRT & RSS)
0.5
SSS (FEC)
SSS (non-FT)
0
2
4
6
8
10
12
Number of Servers
Figure 11.10 Client buffer requirement versus number of servers
Finally, we can observe that only sub-schedule striping under FEC, and under PRT with RSS
are scalable, the latter being completely independent of the system scale. Interestingly, buffer
requirements under FEC decrease for more servers and approach the non-fault-tolerant case.
This is because the level of redundancy is fixed and hence the redundancy overhead incurred
decreases when more servers are added.
11.8 Summary
In this chapter, we have investigated protocols and algorithms to support server-level fault
tolerance in the concurrent push architecture. In particular, we presented and analyzed two
fault-tolerant protocols: FEC and PRT, and two striping policies: block striping and sub-
schedule striping. The first result is that FEC is simpler in implementation, does not require
failure detection, and inherently scalable to any number of servers under sub-schedule strip-
ing. The only downside is additional network bandwidth overhead during normal operation.
Surprisingly, analytical results show that PRT is not scalable if redundant data is distributed
over all servers (similar to RAID-5 in disk arrays), even with sub-schedule striping. To tackle
this problem, we propose storing redundant data centrally in redundant servers to avoid the
reconfiguration delay. We increase the buffer holding time at the redundant servers to enable
Search WWH ::




Custom Search