Takashi Suzuoka

Impact of Job Mix on Optimizations for Space Sharing Schedulers

Abstract Modern parallel systems with N nodes can concurrently service multiple jobs requesting a total of up to to N nodes. One of the challenges for the operating system is to give reasonable service to a diverse group of jobs. Asequence of large jobs, each requiring over half of the available nodes, can reduce the machine utilization by up to 50%, but scheduling a long running job on the idle nodes may block the stream of large jobs. Various policies have been proposed for scheduling parallel computers, but as the users of current supercomputers know, these policies are far from perfect. This paper reports on the measurement of the usage of a 512-node IBM SP2 at Cornell Theory Center, a 96-node Intel Paragon at ETH Zurich, and a 512-node Cray T3D at Pittsburgh Supercomputing Center. We discuss the characteristics of the different workloads and examine their impact on job scheduling. We specifically show how two simple scheduling optimizations based on reordering the waiting queue can be used effectively to improve scheduling performance on real workloads. Supercomputer workloads from different installations exhibit some common characteristics, but they also differ in important ways We demonstrate how this knowledge can be exploited in the design and tuning of schedulers.

Implementation of the prodigy parallel AI machine and performance evaluation of communication

At present, the binary n-cube is considered to be interesting as an interconnection network for the highly parallel computers. Most of those networks, however, use the serial communication channel, which prevents the realization of satisfactory communication performance. From such a viewpoint, it is desired to construct an interconnection network with higher speed, higher versatility, and the easy implementation. The authors have already proposed the base-m n-cube as the interconnection network for the parallel AI machine Prodigy. This paper shows that the base-m n-cube is better than the binary n-cube in versatility and easy implementation. Furthermore, the authors developed Prodigy which utilizes the forementioned excellent property to connect 512 processors by the 8-bit width full duplex base-8 3-cube with a handshake. Two kinds of gate arrays are used in the implementation. It is shown as a result that the interconnection network can be constructed with the highest speed in the presently known cube-type highly parallel computers. It is verified by an actual measurement that the transfer performance is deteriorated only slightly by the collision in the random communication.

Network Performance of the Prodigy Parallel AI Machine

This paper analyzes the network performance of the Prodigy parallel AI machine, which is currently being developed. This machine is a multiprocessor system focusing on knowledge information processing problems with fine-grain parallelism, and thus it is adequate for solving the problems whose structures can be treated as networks. A base-m n-cube network is proposed as an appropriate interconnection network for these problems and the proposed network is employed for the Prodigy interconnection network. It deals with the suitability of the base-m n-cube network to the Prodigy machine and the possibility of making a large system in addition to network performance through comparisons with binary n-cube and mesh networks, where the performance figures are measured by simulation. The results of the comparison show that the base-m n-cube network is better than others for the Prodigy machine.