Steven W. Hunter

Workload and network-optimized computing systems

This paper describes a recent system-level trend toward the use of massive on-chip parallelism combined with efficient hardware accelerators and integrated networking to enable new classes of applications and computing-systems functionality. This system transition is driven by semiconductor physics and emerging network-application requirements. In contrast to general-purpose approaches, workload and network-optimized computing provides significant cost, performance, and power advantages relative to historical frequency-scaling approaches in a serial computational model. We highlight the advantages of on-chip network optimization that enables efficient computation and new services at the network edge of the data center. Software and application development challenges are presented, and a service-oriented architecture application example is shown that characterizes the power and performance advantages for these systems. We also discuss a roadmap for next-generation systems that proportionally scale with future networking bandwidth growth rates and employ 3-D chip integration methods for design flexibility and modularity.

Workload- based power management for parallel computer systems

This paper describes and evaluates predictive power management algorithms that we have developed to minimize energy consumption and unmet demand in parallel computer systems. The algorithms are evaluated using workload data obtained from production servers from several applications, showing that energy savings of 20% or more can readily be achieved, with a small degree of unmet demand and acceptable reliability, availability, and serviceability (RAS) impact. The implementation of these algorithms in IBM system management software and the possibilities for future work are discussed.

Applying Amdahl's other law to the data center

As computing system workloads become more distributed in nature, there is an increasing dependence on the networking interconnects between such systems. As stated by Amdahls Other Law, this dependence not only exists on the I/O (input/output) subsystem, but also on the memory subsystems. In particular, as processor utilization increases, there is a direct, corresponding increase in memory and I/O utilization. At a broader level, the distribution of workloads is driving the need for computing based on locality (or pods) to achieve the appropriate balance of compute, network, and storage resources. This paper studies the applicability of Amdahls Other Law to the data center to better understand the relationship between processor systems and the networks interconnecting them. This study is also relevant as multicore systems will become more prevalent to sustain growth of processing performance.

Combined performance and availability analysis of a switched network application

As switched networks providing services to end users become more commonplace, the integral components of these networks must have an increasing level of dependability. One area of interest is determining the optimal number of network servers required for a switched network application by taking into consideration the characteristics of the network traffic and failure rate of the servers. A stochastic reward net model is developed for this purpose. Such models allow performance and availability measures to be combined.

GPFS-based implementation of a hyperconverged system for software defined infrastructure

The need for an increasingly dynamic and more cost-efficient data-center infrastructure has led to the adoption of a software defined model that is characterized by: the creation of a federated control plane to judiciously allocate and control appropriate heterogeneous infrastructure resources in an automated fashion, the ability for applications to specify criteria, such as performance, capacity, and service levels, without detailed knowledge of the underlying infrastructure; and the migration of data-plane capabilities previously embodied as purpose-built devices or firmware into software running on a standard operating systems in commercial off-the-shelf servers. This last trend of hardware-based capabilities migrating to software is enabling yet another shift to hyperconvergence, which refers to merger of traditionally separate networking, compute, and storage capabilities in integrated system software. This paper examines the convergence of the software defined infrastructure stack, and introduces a hyperconverged compute and storage architecture, in which the IBM General Parallel File System (GPFS®) implements the software defined data plane that dynamically supports workloads ranging from high-I/O virtual desktop infrastructure applications to more compute-oriented analytics applications. The performance and scalability characteristics of this architecture are evaluated with a prototype implementation.

High-performing scale-out solution for deep packet processing via adaptive load-balancing

We propose a scale-out solution for deep packet processing (DPP) appliances, which uses a standard Ethernet switch in combination with a load balancing controller. The majority of the data-plane traffic is distributed via the switchs built-in traffic distribution function and no connection state is kept. If the load balancing controller detects a skew in the load of the DPP appliances, it updates the traffic rules in the switch to redirect new connections to less busy DPP appliances. This adaptive solution is beneficial for load-balancing at high data rates because state is kept for only the redirected connections. For typical traffic patterns, our solution reduces the packet drops with minimal connection redirection. We show an example capture, where we are able to improve the packet drop rate by 94.8% with only 3% of the connections being redirected.

The effects of LAN-like input traffic on a switch access node

This paper addresses the modeling and analysis of a switching network access node. This is an area of importance due to forthcoming high-speed ATM backbone networks and the need to transmit data for current and future network applications. The nature of these applications must be considered when designing an access node to a switching network. A traffic arrival process for the network and characteristics of the access node are combined into an access node performance model. This model is implemented using Deterministic and Stochastic Petri Nets (DSPNs) which provides transitions with zero and deterministic firing times, as well as, exponentially distributed firing times, thus allowing the slotted behavior of the ATM interface to be captured in the model.