E. Rosti, G. Serazzi, E. Smirni, and M. S. Squillante. Models of parallel applications with large computation and I/O requirements. IEEE Trans. on Software Engineering, 28:286-- 307, March 2002.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....can run. At the same time, the effect on the disk performance is the opposite: I O bound processes keep the disks busy, while compute bound processes leave them idle. Indeed, it is non trivial to balance the use of these resources in applications that have large computation and I O requirements [15]. The core idea of gang scheduling is assigning as many processors to a job as are required at the same time. Such an assignment allows the job to avoid blocking processes while they wait for the completion of communications with other processes, because of two reasons: 1. It is guaranteed that ....

....job, whereas setting G comp = 0 creates an I O bound job. Making G comp and or G I=O large increases the granularity (the amount of computation or I O between barriers) This structure of alternating compute and I O phases is widely accepted as a reasonable model for parallel applications [15]. The barrier is implemented by all processes sending a message to process 0, and waiting for a reply. Process 0 sends the reply only after it receives a message 9 loop N times compute part loop G comp times null statement I O part loop G I=O times open new file write B bytes close ....

[Article contains additional citation context not shown here]

Rosti E., Serazzi G., Smirni E., and Squillante M. S., Models of parallel applications with large computation and I/O requirements. IEEE Trans. on Software Engineering 28(3), pp. 286--307, Mar 2002.

....can run. At the same time, the effect on the disk performance is the opposite: I Obound processes keep the disks busy, while compute bound processes leave them idle. Indeed, it is non trivial to balance the use of these resources in applications that have large computation and I O requirements [2]. The core idea of gang scheduling is assigning as many processors to a job as are required at the same time. Such an assignment allows the job to avoid blocking processes while they wait for the completion of communications with other processes, because of two reasons: 1. It is guaranteed that ....

.... job, whereas setting G comp = 0 creates an I O bound job Making G comp and or G I=O large increases the granularity (the amount of computation or I O between barriers) This structure of alternating compute and I O phases is widely accepted as a reasonable model for parallel applications [2]. The barrier is implemented by all processes sending a message to process 0, and waiting for a reply. Process 0 sends the reply only after it receives a message from every other process in the job. In the following results, we use a mix of compute bound and I O bound jobs, with G comp = 2; 500; ....

[Article contains additional citation context not shown here]

Rosti E., Serazzi G., Smirni E., and Squillante M. S., Models of Parallel Applications with Large Computation and I/O Requirements. IEEE Trans. on Software Engineering, Mar 2002.

....average size of work units and its variability, the way in which these work units are partitioned into threads, and the number of barriers by which they are synchronized [30] The question is what distributions to use. While there has been some work done on characterizing speci c applications [20, 74, 65], there has been little if any work on characterizing the mix of application characteristics in a typical workload. A rather singular example is the Charisma project, in which a whole workload was measured [59] Interestingly, this requires the same statistical techniques described in Section 3.2, ....

E. Rosti, G. Serazzi, E. Smirni, and M. S. Squillante, \Models of parallel applications with large computation and I/O requirements". IEEE Trans. Softw. Eng. 28(3), pp. 286-307, Mar 2002.

....run. At the same time, the in uence on the disk performance is the opposite: I O bound processes keep the disk busy, while compute bound processes keep the disk idle. Indeed, it is non trivial to balance the use of these resources in applications that have large computation and I O requirements [2]. If the characteristics of each gang are known, this can be exploited in order to keep both the CPU and the I O devices busy. This proposal presents the idea of matching pairs of gangs, one compute bound and the other I O bound. The rationale for such matching is that these gangs will hardly ....

....Note, however, that the implementation of the barrier using an external server creates some I O activity even for compute bound 8 Figure 5: Barriers per second. processes. This structure of alternating compute and I O phases is widely accepted as a reasonable model for parallel applications [2]. The experiments were conducted with a static job mix. In other words, a set of jobs were all started at the same time, and each job s size was the same as the cluster size (8 nodes were used unless otherwise noted) Performance data was collected during the interval in which all the jobs ran ....

[Article contains additional citation context not shown here]

Rosti E., Serazzi G., Smirni E., and Squillante M. S., Models of Parallel Applications with Large Computation and I/O Requirements. IEEE Trans. on Software Engineering, to appear, 2001.

No context found.

E. Rosti, G. Serazzi, E. Smirni, and M. S. Squillante. Models of parallel applications with large computation and I/O requirements. IEEE Trans. on Software Engineering, 28:286-- 307, March 2002.

No context found.

E. Rosti, G. Serazzi, E. Smirni, and M.S. Squillante, "Models of parallel applications with large computation and I/O requirements," IEEE Transactions on Software Engineering, vol. 28(3), pp. 286--307, 2002.

No context found.

E. Rosti, G. Serazzi, E. Smirni, and M. S. Squillante. Models of parallel applications with large computation and I/O requirements. IEEE Trans. on Software Engineering, 28:286-- 307, March 2002.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC