Programming languages that provide multidimensional arrays and a flat linear model of memory must implement a mapping between these two domains to order array elements in memory. This layout function is fixed at language definition time and constitutes an invisible, non-programmable array attribute. In reality, modern memory systems are architecturally hierarchical rather than flat, with substantial differences in performance among different levels of the hierarchy. This mismatch between the model and the true architecture of memory systems can result in low locality of reference and poor performance. Some of this loss in performance can be recovered by re-ordering computations using transformations such as loop tiling. We explore nonlinear array layout functions as an additional means of improving locality of reference. For a benchmark suite composed of dense matrix kernels, we show by timing and simulation that two specific layouts (4D and Morton) have low implementation costs (2-5% of total running time) and high performance benefits (reducing execution time by factors of 1.1-2.5); that they have smooth performance curves, both across a wide range of problem sizes and over representative cache architectures; and that recursion-based control structures may be needed to fully exploit their potential.

The role of the operating system (OS) in managing shared resources such as CPU time, memory, peripherals, and even energy is well motivated and understood [23]. Unfortunately, one key resource-lower-level shared cache in chip multi-processors-is commonly managed purely in hardware by rudimentary replacement policies such as least-recentlyused (LRU). The rigid nature of the hardware cache management policy poses a serious problem since there is no single best cache management policy across all sharing scenarios. For example, the cache management policy for a scenario where applications from a single organization are running under "best effort" performance expectation is likely to be different from the policy for a scenario where applications from competing business entities (say, at a third party data center) are running under a minimum service level expectation. When it comes to managing shared caches, there is an inherent tension between flexibility and performance. On one hand, managing the shared cache in the OS offers immense policy flexibility since it may be implemented in software. Unfortunately, it is prohibitively expensive in terms of performance for the OS to be involved in managing temporally fine-grain events such as cache allocation. On the other hand, sophisticated hardware-only cache management techniques to achieve fair sharing or throughput maximization have been proposed. But they offer no policy flexibility. This paper addresses this problem by designing architectural support for OS to efficiently manage shared caches with a wide variety of policies. Our scheme consists of a hardware cache quota management mechanism, an OS interface and a set of OS level quota orchestration policies. The hardware mechanism guarantees that OS-specified quotas are enforced in shared caches, thus eliminating the need for (and the performance penalty of) temporally fine-grained OS intervention. The OS retains policy flexibility since it can tune the quotas during regularly scheduled OS interventions. We demonstrate that our scheme can support a wide range of policies including policies that provide (a) passive per-

Minimizing latency and maximizing throughput are important goals in the design of routing algorithms for interconnection networks. Ideally, we would like a routing algorithm to (a) route packets using the minimal number of hops to reduce latency and preserve communication locality, (b) deliver good worst-case and averagecase throughput and (c) enable low-complexity (and hence, low latency) router implementation. In this paper, we focus on routing algorithms for an important class of interconnection networks: two dimensional (2D) mesh networks. Existing routing algorithms for mesh networks fail to satisfy one or more of design goals mentioned above. Variously, the routing algorithms suffer from poor worst case throughput (ROMM [13], DOR [24]), poor latency due to increased packet hops (VALIANT [32]) or increased latency due to hardware complexity (minimal-adaptive [7,31]). * This report is an expanded version of a previously published paper [18]. 1The major contribution of this paper is the design of an oblivious routing algorithm-O1TURN-with provable near-optimal worst-case throughput, good average-case throughput, low design complexity and minimal number of network hops for 2D-mesh networks, thus satisfying all the stated design goals. O1TURN offers optimal worst-case throughput when the network radix (k in a kxk network) is even. When the network radix is odd, O1TURN is within a 1/k 2 factor of optimal worst-case throughput. O1TURN achieves superior or comparable average-case throughput with global traffic as well as local traffic. For example, O1TURN achieves 18.8%, 0.7% and 13.6% higher average-case throughput than DOR, ROMM and VALIANT routing, respectively when averaged over one million random traffic patterns on an 8x8 network. Finally, we demonstrate that O1TURN is well suited for a partitioned router implementation that is of similar delay complexity as a simple dimension-ordered router. Our implementation incurs a marginal increase in switch arbitration delay that is completely hidden in pipelined routers as it is not on the clock-critical path.

No abstract

Network performance in tightly-coupled multiprocessors typically degrades rapidly beyond network saturation. Consequently, designers must keep a network below its saturation point by reducing the load on the network. Congestion control via source throttling-a common technique to reduce the network load-prevents new packets from entering the network in the presence of congestion. Unfortunately, prior schemes to implement source throttling either lack vital global information about the network to make the correct decision (whether to throttle or not) or depend on specific network parameters, network topology, or communication patterns. This paper presents a global-knowledge-based, selftuned, congestion control technique that prevents saturation at high loads across different network configurations and communication patterns. Our design is composed of two key components. First, we use global information about a network to obtain a timely estimate of network congestion. We compare this estimate to a threshold value to determine when to throttle packet injection. The second component is a self-tuning mechanism that automatically determines appropriate threshold values based on throughput feedback. A combination of these two techniques provides high performance under heavy load, does not penalize performance under light load, and gracefully adapts to changes in communication patterns.