Vahid Tabatabaee

Parallel Parameter Tuning for Applications with Performance Variability

In this paper, we present parallel on-line optimization algorithms for parameter tuning of parallel programs. We employ direct search algorithms that update parameters based on real-time performance measurements. We discuss the impact of performance variability on the accuracy and efficiency of the optimization algorithms and proposed modified versions of the direct search algorithms to cope with it. The modified version uses multiple samples instead of single sample to estimate the performance more accurately. We present preliminary results that the performance variability of applications on clusters is heavy tailed. Finally, we studay and demonstrate the performance of the proposed algorithms for real scientific application.

The Network Processing Forum switch fabric benchmark specifications: an overview

The Network Processing Forum chartered a fabric benchmarking task group to establish a set of switch fabric benchmark specifications that allows the characterization of a wide range of switch fabrics for diverse networking applications. A unique characteristic of the benchmarks is their ability to produce comparable performance results for different switch fabrics, regardless of their underlying architecture and technology. This article provides an overview of the NPF fabric benchmark specifications by describing the various topics addressed by the standard as well as their potential impact.

Tuning parallel applications in parallel

In this paper, we present and evaluate a parallel algorithm for parameter tuning of parallel applications. We discuss the impact of performance variability on the accuracy and efficiency of the optimization algorithm and propose a strategy to minimize the impact of this variability. We evaluate our algorithm within the Active Harmony system, an automated online/offline tuning framework. We study its performance on three benchmark codes: PSTSWM, HPL and POP. Compared to the Nelder-Mead algorithm, our algorithm finds better configurations up to seven times faster. For POP, we were able to improve the performance of a production sized run by 59%.

QoS provisioning and tracking fluid policies in input queueing switches

The concept of tracking fluid policies by packetized policies is extended to input queueing switches. It is considered that the speedup of the switch is one. One of the interesting applications of the tracking policy in TDMA satellite switches is elaborated. For the special case of 2 × 2 switches, it is shown that a tracking nonanticipative policy always exists. It is found that, in general, nonanticipative policies do not exist for switches with more than two input and output ports. For the general case of N × N switches, a heuristic tracking policy is provided. The heuristic algorithm is based on two notions: port tracking and critical links. These notions can be employed in the derivation of other heuristic tracking policies as well. Simulation results show the usefulness of the heuristic algorithm and the two basic concepts it relies on.

MNCM a new class of efficient scheduling algorithms for input-buffered switches with no speedup

In this paper, we use fluid model techniques to establish some new results for the throughput of input-buffered switches. In particular, we introduce a new class of deterministic maximal size matching algorithms that achieves 100% throughput. Dai and Prabhakar (2000) has shown that any maximal size matching algorithm with speedup of 2 achieves 100% throughput. We introduce a class of maximal size matching algorithms that we call them maximum node containing matching (MNCM) algorithms, and prove that they have 100% throughput with no speedup. We also introduce a new weighted matching algorithm, maximum first matching (MFM) with complexity O(N2.5) that belongs to MNCM. MFM, to the best of our knowledge, is the lowest complexity deterministic algorithm that delivers 100% throughput. The only assumption on the input traffic is that it satisfies the strong law of large numbers. Besides throughput, average delay is the other key performance metric for the input-buffered schedulers. We use simulation results to compare and study the delay performance of MFM. The simulation results demonstrate promising delay performance for MFM.

Robust Routing with Unknown Traffic Matrices

In this paper, we present an algorithm for intra-domain traffic engineering. We assume that the traffic matrix, which specifies traffic load between every source-destination pair in the network, is unknown and varies with time, but that always lies inside an explicitly defined region. Our goal is to compute a fixed robust routing with best worst case performance for all traffic matrices inside the bounding region. We formulate this problem as a semi-infinite programming problem. Then, we focus on a special case with practical merits, where (1) the traffic matrix region is assumed to be a polytope specified by a finite set of linear inequalities, and (2) our objective is to find the routing that minimizes the maximum link utilization. Under these assumptions, the problem can be formulated as a polynomial size linear programming (LP) problem with finite number of constraints. We further consider two specific set of constraints for the traffic matrix region. The first set is based on the hose model and limits the total traffic rate of network point of presence (PoP) nodes. The second set is based on the pipe model and limits the traffic between source-destination pairs. We study the effectiveness of each set of constraints using extensive simulations.

Automatic software interference detection in parallel applications

We present an automated software interference detection methodology for Single Program, Multiple Data (SPMD) parallel applications. Interference comes from the system and unexpected processes. If not detected and corrected such interference may result in performance degradation. Our goal is to provide a reliable metric for software interference that can be used in soft-failure protection and recovery systems. A unique feature of our algorithm is that we measure the relative timing of application events (i.e. time between MPI calls) rather than system level events such as CPU utilization. This approach lets our system automatically accommodate natural variations in an applications utilization of resources. We use performance irregularities and degradation as signs of software interference. However, instead of relying on temporal changes in performance, our system detects spatial performance degradation across multiple processors. We also include a case study that demonstrates our techniques effectiveness, resilience and robustness.

Modelling and optimization for multi-hop wireless networks using fixed point and automatic differentiation

We develop and evaluate a new method for estimating and optimizing various performance metrics of multi-hop wireless networks, including MANETs. We introduce a simple approximate (throughput) loss model that couples the physical, MAC and routing layers effects. The model provides quantitative statistical relations between the loss parameters that are used to characterize multiuser interference and physical path conditions on the one hand and the traffic rates between origin-destination pairs on the other. The model considers effects of the hidden nodes, node scheduling algorithms, MAC and PHY layer failures and unsuccessful packet transmission attempts at the MAC layer in arbitrary network topologies where multiple paths share nodes. We apply automatic differentiation (AD) to these implicit performance models, and develop a methodology for sensitivity analysis and parameter optimization for wireless protocols. Finally, we provide simulation experiments to evaluate the effectiveness and performance estimation accuracy of the proposed models and methodologies.

Component Based Performance Modelling of Wireless Routing Protocols

We propose a component based methodology for modelling and design of wireless routing protocols. Componen-tization is a standard methodology for analysis and synthesis of complex systems, or software. The feasibility of the component based design relies heavily on the compositionality property (i.e. system-level properties can be computed from properties of components). To provide a component based design methodology and to test compositionality for routing protocols, we have to develop a component based model of the wireless network. We present the main components of the routing protocol that should be modelled and focus on three main components: neighborhood discovery, selector of topology information to disseminate, and the path selection components. For each component, we identify the inputs, outputs, and a generic methodology for modelling. Throughout the paper, we use the Optimized Link State Routing (OLSR) protocol as a case study to demonstrate the effectiveness of our approach. Using the neighborhood discovery component, we present our design methodology and design a modified enhanced version of this component, and compare its performance to the original OLSR design.

MNCM: a critical node matching approach to scheduling for input buffered switches with no speedup

In this paper, we use fluid model techniques to es-tablish new results for the throughput of input-buffered switches. Dai and Prabhakar have shown that any maximal size matching algorithm with speedup of 2 achieves 100% throughput. We introduce the maximum node containing matching (MNCM), which is a new class of matching algorithms that achieve 100% throughput with no speedup. The only assumption on the arrival processes is they satisfy the strong law of large numbers (SLLN). The MNCM policies only need to include ports whose weight (backlog) are above a threshold in the matching rather than finding a matching with maximum total weight. This simplified requirement enables us to introduce a new matching algorithm, maximum first matching (MFM), with O(N2.5) complexity. We show that MFM is a low-complexity algorithm with good delay performance. We also provide a deterministic upper bound for the buffering requirement of a switch with an MNCM scheduler, when the ports incoming traffic are admissible and (σ, ρ) regulated.