Academia.eduAcademia.edu

Figure 9 – uploaded by Fan Yang

See full PDFdownloadDownload figure
Figure 11. Throughput and goodput over time as Sonora checkpoints twice and then recovers from a failure.

Figure 11 Throughput and goodput over time as Sonora checkpoints twice and then recovers from a failure.

Related Figures (11)

Figure 3. A data-flow view of Sonora. handled transparently if this is desirable; an application may also decide to be notified when disconnections oc- cur.
Figure 3. A data-flow view of Sonora. handled transparently if this is desirable; an application may also decide to be notified when disconnections oc- cur.
Figure 4. Sonora system architecture.
Figure 4. Sonora system architecture.
Figure 5. An example of failure recovery. vo s Ulputl VCIUCEsS. More precisely, let F’ be the set of vertices that have failed and Dr be the set of downstream vertices from any vertex in fF’. Then the rollback set Ry is the union of F and Dp: these vertices will be rolled back to the latest consistent checkpoint. The re-execution set Ep contains all input vertices of Rr that are not in Rp. Tc restart vertices in Ry from the checkpoint, vertices in Er must replay inputs. If the recorded output for re- execution is lost on a vertex v in Ep, vertex v has to be added to the rollback set and the re-execution has to start from an earlier stage. In the worst case, all ver- tices are in the rollback set and the entire computation restarts from that latest consistent checkpoint. Correct- ness of recovery follows directly from global consis- tency provided by the global checkpointing protocol, as a special case of the Chandy-Lamport’s snapshot pro- tocol. pany a cag wees es cae — a
Figure 5. An example of failure recovery. vo s Ulputl VCIUCEsS. More precisely, let F’ be the set of vertices that have failed and Dr be the set of downstream vertices from any vertex in fF’. Then the rollback set Ry is the union of F and Dp: these vertices will be rolled back to the latest consistent checkpoint. The re-execution set Ep contains all input vertices of Rr that are not in Rp. Tc restart vertices in Ry from the checkpoint, vertices in Er must replay inputs. If the recorded output for re- execution is lost on a vertex v in Ep, vertex v has to be added to the rollback set and the re-execution has to start from an earlier stage. In the worst case, all ver- tices are in the rollback set and the entire computation restarts from that latest consistent checkpoint. Correct- ness of recovery follows directly from global consis- tency provided by the global checkpointing protocol, as a special case of the Chandy-Lamport’s snapshot pro- tocol. pany a cag wees es cae — a
tem catches up with the load. Redirec take advantage of periods of low load ion helps Sonora for catching up, and enables batching for higher throughput. In case the incoming data rate exceeds the t hroughput of the storage system, the cloud pushes back onto the mobile devices to adaptively reduce the sensor sampling fre- quency (see Figure 2 and Section 2.1 for details). storage system, the cloud pushes back onto the mobile
tem catches up with the load. Redirec take advantage of periods of low load ion helps Sonora for catching up, and enables batching for higher throughput. In case the incoming data rate exceeds the t hroughput of the storage system, the cloud pushes back onto the mobile devices to adaptively reduce the sensor sampling fre- quency (see Figure 2 and Section 2.1 for details). storage system, the cloud pushes back onto the mobile
Figure 7. PEIR is CPU-bounded.
Figure 7. PEIR is CPU-bounded.
Figure 8. Sonora scalability. Figure 8 shows that the Sonora implementation of PEIR scales linearly — peak throughput increases pro- portionally to the number of machines. The speedup efficiency is 0.70 when the number of machines is 40. If each mobile user reported their location once every 5 seconds, Sonora would be able to support over 350,000 users concurrently with just 40 machines. In addition, sync stream filters can reduce the amount of traffic from mobile devices by filtering out insignificant changes in location. This would allow Sonora to support even more users.
Figure 8. Sonora scalability. Figure 8 shows that the Sonora implementation of PEIR scales linearly — peak throughput increases pro- portionally to the number of machines. The speedup efficiency is 0.70 when the number of machines is 40. If each mobile user reported their location once every 5 seconds, Sonora would be able to support over 350,000 users concurrently with just 40 machines. In addition, sync stream filters can reduce the amount of traffic from mobile devices by filtering out insignificant changes in location. This would allow Sonora to support even more users.
Figure 9. Online load balancing.
Figure 9. Online load balancing.
Figure 10. CPU load distribution. were overloaded because there were no additional idle machine (with CPU load below 30%) to share the load. Fault tolerance. To evaluate how Sonora recovers from machine failures we turned off dynamic adap- tation and ran PEIR on 32 machines, with 9 machines running a physical vertex in the first stage and 23 ma- chines running a physical vertex in the second stage. This assignment is balanced as the second stage is more costly than the first. All machines in the first stage generate data that is consumed by all machines in the second stage. A dispatcher was used to feed input to vertices in the first stage. were overloaded because there were no additional idle
Figure 10. CPU load distribution. were overloaded because there were no additional idle machine (with CPU load below 30%) to share the load. Fault tolerance. To evaluate how Sonora recovers from machine failures we turned off dynamic adap- tation and ran PEIR on 32 machines, with 9 machines running a physical vertex in the first stage and 23 ma- chines running a physical vertex in the second stage. This assignment is balanced as the second stage is more costly than the first. All machines in the first stage generate data that is consumed by all machines in the second stage. A dispatcher was used to feed input to vertices in the first stage. were overloaded because there were no additional idle
Figure 12. Total KB sent over time with a sync stream. A disconnection is masked between ¢) and
Figure 12. Total KB sent over time with a sync stream. A disconnection is masked between ¢) and
Related topics:
Failure RecoveryRollback recovery
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved