NSF Grants

Project Title : I3C: An Infrastructure for Innovation in Information Computing
Funding Agency : National Science Foundation

The CISE Research Infrastructure grant will be used for setting up a 128-node (64+64) Linux cluster, a 32-node Sun cluster, a storage area network (SAN) and a wireless infrastructure for supporting coordinated research of 25 investigators. The multidisciplinary research that spans across many engineering and science disciplines focuses on three core areas that are essential to advance the state-of-the-art in cluster computing. The first area is applications, which is the main driving force of the infrastructure. Here our strength lies in three specific application domains; Computational Science, Digital Immortality, and Bioinformatics. In addition, the clusters will be used extensively for simulation of various complex systems. The second area of research is System Software, which examines how various cluster resources can be effectively used for improving the delivered performance. Finally, we investigate many low-level Architectural issues that are essential to provide high and assured performance. The research is proposed in a cohesive manner explaining the interaction among the three main areas.

Project Title : An Integrated Approach for Quality of Service in Cluster Networks
Funding Agency : National Science Foundation

With the increasing use of cluster systems for a variety of interactive applications, predictable communication performance or Quality of Service (QoS), has become a major concern. The motivation of this research was to design a cluster communication infrastructure to support different applications with varying QoS requirements. In particular, the research focused on  using the wormhole switching paradigm, which has been used in designing many commercial routers such as the Marcum's Myrinet and IBM SP2, for providing QoS guarantees. The research activities include (1) wormhole router architecture design and evaluation to provide predictable and high performance; (2) network interface and software messaging layer support to inject/eject traffic into/from the network as per the QoS requirements; and (3) CPU scheduling mechanisms that propagate these capabilities up to the application level.

Two main contributions of this research are the following: We have shown how the commercially successful wormhole routers can be extended with minimal hardware modifications to support QoS in clusters. The proposed router, called MediaWorm, includes two modifications to provide predictable performance. First, the virtual channels (VCs), which were originally proposed for improving performance by time-multiplexing different packets/flits on the same physical channel, were statically allocated to different traffic classes. Second, the traditional FCFC or round robin (RR) scheduling was replaced by the VirtualClock scheduling to provide rate proportional bandwidth to different traffic classes. It was shown that a cluster designed with the MediaWorm routers can  provide soft guarantees to MPEG II media streams in the presence of both best-effort and media traffic. Next, the design was extended for allocating the VCs dynamically to different classes of traffic.

Second, for end-to-end QoS assurance, we have designed a QoS-capable NIC based on the virtual interface architecture (VIA) design paradigm. The design involves three modifications to the original VIA: (i) Inclusion of a prioritized doorbell scheme for informing the NIC the arrival of different traffic classes; (ii) Partitioning of the NIC buffer to a number of VCs compatible with the router design; and (iii) Providing a rate proportional scheduling of the VCs to inject flits into the network. Co-evaluation of the QoS-capable routers and QoS-capable NICs revealed that QoS provisioning in the NIC is more critical than that in the router/network. We believe that our NIC study is the first effort in highlighting the importance of network interface design for QoS support.

Project Title :  Scalable and Efficient Scheduling Techniques for Clusters
Funding Agency : National Science Foundation

The main motivation of this research is to design scalable and efficient scheduling algorithms for clusters. The proposed research addresses three closely intertwined issues for developing such algorithms. First, an in-depth evaluation of the existing communication-induced scheduling schemes will be done  using real workloads to investigate their performance, implementation complexity, scalability, and fairness properties. With a better understanding of the strength and weakness of these policies, Second, since these scheduling algorithms rely on a low-latency, user-level communication mechanism, various design issues in implementing these algorithms will be explored. The last component of the research will examine how the scheduling algorithms can be tailored to facilitate predictable performance in clusters. The research is being  conducted in collaboration with the Penn State's Center for Academic Computing (CAC) group and the Lawrence Livermore National Laboratory (LLNL).

We have developed a generic framework for implementing all prior coscheduling on a Linux cluster with minimal overhead. The framework has been implemented on a 16-node Linux cluster connected through Myrinet. Three prior coscheduling techniques (DCS, SB and PB) have been implemented and analyzed using this framework. We have proposed  two new coscheduling algorithms, called Co-ordinated Coscheduling (CC) and Hybrid coscheduling, which can outperform not only all prior coscheduling techniques but also the traditional batch scheduling (PBS). In particular, we believe that the proposed Hybrid scheme has the potential to replace the currently used batch scheduling techniques in clusters. We are currently evaluating all the coscheduling techniques  on large clusters with LLNL and CAC workloads to study the scalability issue.

Project Title  :  QoS Provisioning in InfiniBand Architecture (IBA) for System Area Networks
Funding Agency : National Science Foundation

The InfiniBandTM Architecture (IBA) is envisioned to be the default communication infrastructure for future System Area Networks (SANs) or clusters. The InfiniBand Trade Association (IBTA)  has released the first IBA specification, and is currently augmenting it with enhanced features such as congestion management, Quality of Service (QoS) and router management. However, the IBA design is currently in its infancy since the released specification outlines only the high level functionalities, leaving it open for the research and industrial community to explore various design alternatives. In particular, QoS support in IBA is targeted as a critical issue to support many server applications with real-time constraints. However, design of IBA for QoS is an unexplored area of research and is the main focus of this investigation.

The main objective of this ongoing research is to explore various design alternatives for providing high and predictable performance in IBA-style SANs. It covers the design of an IBA fabric, design of congestion avoidance techniques, developing multicasting algorithms, and developing fault-tolerant techniques. The research is aimed at developing and releasing a complete IBA simulator platform for various types of research on IBA.

We have proposed four different techniques for improving the performance of IBA-style SANs. These include using the Internet compliant Shortest Path First (SPF) routing algorithm in stated of the UP/DOWN routing, a  fault-tolerant version of the SPF routing to provide automatic path migration (APM) in the presence of faults, a packet dropping mechanism to avoid  network congestion, and efficient hardware implementation of multicasting. These are developed based on the IBA specification. We have developed an IBA simulator and is currently augmenting the simulator with additional features. We have also developed an energy model and an energy optimization technique, called dynamic link shutdown (DLS), for the cluster network. It is shown that the DLS scheme can provide better performance-energy  tradeoffs compared to the known dynamic voltage scaling (DVS) scheme. Currently, we are extending our IBA design for better performance and energy saving.