Characterizing the Influence of System Noise on Large-Scale Applications by Simulation
read more
Citations
Scientific benchmarking of parallel computing systems: twelve ways to tell the masses when reporting performance results
There goes the neighborhood: performance degradation due to nearby jobs
Hiding Global Communication Latency in the GMRES Algorithm on Massively Parallel Machines
Using automated performance modeling to find scalability bugs in complex codes
Enabling highly-scalable remote memory access programming with MPI-3 one sided
References
MPI: A Message-Passing Interface Standard
BoomerAMG: a parallel algebraic multigrid solver and preconditioner
The Case of the Missing Supercomputer Performance: Achieving Optimal Performance on the 8,192 Processors of ASCI Q
Characterizing application sensitivity to OS interference using kernel-level noise injection
System noise, OS clock ticks, and fine-grained parallel applications
Related Papers (5)
Frequently Asked Questions (12)
Q2. What are the future works in "Characterizing the influence of system noise on large-scale applications by simulation" ?
A possible direction for future research is the use of nonblocking collective operations that separate starting the operation and waiting for the data.
Q3. What is the main purpose of the simulator?
The simulator supports the injection of system noise into all computations that are performed on the CPU: the application computation, the LogGOPS overheads os, or, and O, and the reductions in collective operations.
Q4. What is the direction for future research?
A possible direction for future research is the use of nonblocking collective operations that separate starting the operation and waiting for the data.
Q5. What did the authors find out about the effect of co-scheduling?
Co-scheduling eliminated nearly all noise propagation and showed excellent scaling behavior so that results were omitted from the graphs (less than 0.5% slowdown).
Q6. Why are faster networks not able to improve the application speed significantly?
at large-scale, faster networks are not able to improve the application speed significantly because noise propagation is becoming a bottleneck.
Q7. What is the synchronization overhead on receiver and rendezvous-sender?
The synchronization overhead on receiver and rendezvous-sender can be modeled as Xr = max{Ts+σTs +N −Twr , 0} andXs = max{Tr + σTr −N −Tws , 0}, respectively.
Q8. How many processes were predicted with the simulator?
The simulator was shown to predict collective operations up to 128 processes with an average error of less than 1% and full MPI applications with an error below 2%.
Q9. Why do the authors expect nonblocking communication to be relatively resistant to system noise?
The authors expect applications that use nonblocking communication to be relatively resistant to system noise due to the possibility to hide some synchronization overheads.
Q10. What are the longest paths in the 15- process example?
The longest paths in the 15- process example are (0, 1, 3, 7), (0, 2, 6, 14), (0, 1, 5, 13), and (0, 1, 3, 11) with four recv/send dependencies along each path.
Q11. How do the authors analyze the effects of noise on Linux?
The authors analyze those effects in detail by utilizing the LogGOPS network model to characterize all situations where noise is transported or absorbed.
Q12. What does the author think of the influence of noise on the network parameters?
5) Influence of the Network Parameters: Several previous studies suggest that the influence of noise is tightly coupled to the network parameters.