My research focuses on high-speed communication in parallel and distributed computer systems and in particular on the integration of communication and computation. Network performance has increased dramatically over the past few years, yet the communication performance seen by most applications has not increased by nearly the same amount. The major reason for this discrepancy lies in the inefficiency of the communication and networking layers used in standard operating systems. My group is currently developing new communication layers for parallel machines, networks of workstations, and heterogeneous combinations of the two.
Several years ago, I developed Active Messages, a new communication mechanism that reduces the cost of communication on existing parallel machines and provides a framework for developing efficient new communication hardware. I am involved in several high-level parallel language projects (Split-C and CC++) that are based on Active Messages, and my group has recently completed a port of Active Messages to the IBM SP-2.
We have been experimenting with ATM networks and have implemented a user-level communication architecture, called U-Net, for our ATM cluster, which consists of SPARCstations and a Fore Systems ATM network. U-Net allows processes to control the ATM network interface directly and thereby circumvents the expensive traditional networking layers in the operating system. The design of U-Net ensures that protection between processes can be maintained. We have ported Active Messages and Split-C to U-Net. First benchmarking efforts indicate that the network of 8 SPARCstations outperforms an 8-node Thinking Machines CM-5 or an 8-node Meiko CS-2.
If you have questions or comments please contact: firstname.lastname@example.org.