Precise hardware performance models play a crucial role in code optimizations. They can assist compilers in making heuristic decisions or aid autotuners in identifying the optimal configuration for a given program. For example, the autotuner for XLA, a machine learning compiler, discovered 10–20\% speedup on state-of-the-art models serving substantial production traffic at Google. Although there exist a few datasets for program performance prediction, they target small sub-programs such as basic blocks or kernels. This paper introduces TpuGraphs, a performance prediction dataset on full tensor programs, represented as computational graphs, running on Tensor Processing Units (TPUs). Each graph in the dataset represents the main computation of a machine learning workload, e.g., a training epoch or an inference step. Each data sample contains a computational graph, a compilation configuration, and the execution time of the graph when compiled with the configuration. The graphs in the dataset are collected from open-source machine learning programs, featuring popular model architectures (e.g., ResNet, EfficientNet, Mask R-CNN, and Transformer). TpuGraphs provides 25x more graphs than the largest graph property prediction dataset (with comparable graph sizes), and 770x larger graphs on average compared to existing performance prediction datasets on machine learning programs. This graph-level prediction task on large graphs introduces new challenges in learning, ranging from scalability, training efficiency, to model quality.

Search-based techniques have been demonstrated effective in solving complex optimization problems that arise in domain-specific compilers for machine learning (ML). Unfortunately, deploying such techniques in production compilers is impeded by two limitations. First, prior works require factorization of a computation graph into smaller subgraphs over which search is applied. This decomposition is not only non-trivial but also significantly limits the scope of optimization. Second, prior works require search to be applied in a single stage in the compilation flow, which does not fit with the multi-stage layered architecture of most production ML compilers. This paper presents Xtat, an autotuner for production ML compilers that can tune both graph-level and subgraph-level optimizations across multiple compilation stages. Xtat applies Xtat-M, a flexible search methodology that defines a search formulation for joint optimizations by accurately modeling the interactions between different compiler passes. Xtat tunes tensor layouts, operator fusion decisions, tile sizes, and code generation parameters in XLA, a production ML compiler, using various search strategies. In an evaluation across 150 ML training and inference models on Tensor Processing Units (TPUs) at Google, Xtat offers up to 2.4x and an average 5% execution time speedup over the heavily-optimized XLA compiler.

Many recent machine learning models rely on fine-grained dynamic control flow for training and inference. In particular, models based on recurrent neural networks and on reinforcement learning depend on recurrence relations, data-dependent conditional execution, and other features that call for dynamic control flow. These applications benefit from the ability to make rapid control-flow decisions across a set of computing devices in a distributed system. For performance, scalability, and expressiveness, a machine learning system must support dynamic control flow in distributed and heterogeneous environments.

Bigtable is a distributed storage system for managing structured data that is designed to scale to a very large size: petabytes of data across thousands of commodity servers. Many projects at Google store data in Bigtable, including web indexing, Google Earth, and Google Finance. These applications place very different demands on Bigtable, both in terms of data size (from URLs to web pages to satellite imagery) and latency requirements (from backend bulk processing to real-time data serving). Despite these varied demands, Bigtable has successfully provided a flexible, high-performance solution for all of these Google products. In this article, we describe the simple data model provided by Bigtable, which gives clients dynamic control over data layout and format, and we describe the design and implementation of Bigtable.

A resource may be abused if its users incur little or no cost. For example, e-mail abuse is rampant because sending an e-mail has negligible cost for the sender. It has been suggested that such abuse may be discouraged by introducing an artificial cost in the form of a moderately expensive computation. Thus, the sender of an e-mail might be required to pay by computing for a few seconds before the e-mail is accepted. Unfortunately, because of sharp disparities across computer systems, this approach may be ineffective against malicious users with high-end systems, prohibitively slow for legitimate users with low-end systems, or both. Starting from this observation, we research moderately hard functions that most recent systems will evaluate at about the same speed. For this purpose, we rely on memory-bound computations. We describe and analyze a family of moderately hard, memory-bound functions, and we explain how to use them for protecting against abuses.

Abstract Concurrent programs can suffer from many types of errors, not just the well‐studied problems of deadlocks and simple race conditions on variables. This paper addresses a kind of race condition that arises from reading a variable whose value is possibly out of date. The paper introduces a simple technique for detecting such stale values, and reports on the encouraging experience with a compile‐time checker that uses the technique. Copyright © 2004 John Wiley & Sons, Ltd.

This paper presents novel sampling-based techniques for collecting statistical profiles of register contents, data values, and other information associated with instructions, such as memory latencies. Values of interest are sampled in response to periodic interrupts. The resulting value profiles can be analyzed by programmers and optimizers to improve the performance of production uniprocessor and multiprocessor systems.Our value sampling system extends the DCPI continuous profiling infrastructure, and inherits many of its desirable properties: our value profiler has low overhead (approximately 10% slowdown); it profiles all the code in the system, including the operating system kernel; and it operates transparently, without requiring any modifications to the profiled code.

Article Eraser: a dynamic data race detector for multi-threaded programs Share on Authors: Stefan Savage Department of Computer Science and Engineering, University of Washington, Seattle Department of Computer Science and Engineering, University of Washington, SeattleView Profile , Michael Burrows Digital Equipment Corporation, Systems Research Center Digital Equipment Corporation, Systems Research CenterView Profile , Greg Nelson Digital Equipment Corporation, Systems Research Center Digital Equipment Corporation, Systems Research CenterView Profile , Patrick Sobalvarro Digital Equipment Corporation, Systems Research Center Digital Equipment Corporation, Systems Research CenterView Profile , Thomas Anderson Computer Science Division, University of California, Berkeley Computer Science Division, University of California, BerkeleyView Profile Authors Info & Claims SOSP '97: Proceedings of the sixteenth ACM symposium on Operating systems principlesOctober 1997 Pages 27–37https://doi.org/10.1145/268998.266641Published:01 October 1997 218citation1,521DownloadsMetricsTotal Citations218Total Downloads1,521Last 12 Months6Last 6 weeks0 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

Multithreaded programming is difficult and error prone. It is easy to make a mistake in synchronization that produces a data race, yet it can be extremely hard to locate this mistake during debugging. This article describes a new tool, called Eraser, for dynamically detecting data races in lock-based multithreaded programs. Eraser uses binary rewriting techniques to monitor every shared-monory reference and verify that consistent locking behavior is observed. We present several case studies, including undergraduate coursework and a multithreaded Web search engine, that demonstrate the effectiveness of this approach.

We describe a design for security in a distributed system and its implementation. In our design, applications gain access to security services through a narrow interface. This interface provides a notion of identity that includes simple principals, groups, roles, and delegations. A new operating system component manages principals, credentials, and secure channels. It checks credentials according to the formal rules of a logic of authentication. Our implementation is efficient enough to support a substantial user community.

We describe a design and implementation of security for a distributed system. In our system, applications access security services through a narrow interface. This interface provides a notion of identity that includes simple principals, groups, roles, and delegations. A new operating system component manages principals, credentials, and secure channels. It checks credentials according to the formal rules of a logic of authentication. Our implementation is efficient enough to support a substantial user community.

We study some of the concepts, protocols, and algorithms for access control in distributed systems, from a logical perspective. We account for how a principal may come to believe that another principal is making a request, either on his own or on someone else's behalf. We also provide a logical language for accesss control lists and theories for deciding whether requests should be granted.

We have incorporated on-line data compression into the low levels of a log-structured file system (Rosenblum's Sprite IFS).Each block of data or meta-data is compressed as it is written to the disk and decompressed as it is read.The log-structuring overcomes the problems of allocation and fragmentation for variable-sized blocks.We observe compression factors ranging from 1.6 to 2.2, using algorithms running from 1.7 to 0.4 MBytes per second in software on a DECstation 5000/200.System performance is degraded by a few percent for normal activities (such as compiling or editing), and as much as a factor of 1.6 for file system intensive operations (such as copying multi-megabyte files).Hardware compression devices mesh well with this design.Chips are already available that operate at speeds exceeding disk transfer rates, which indicates that hardware compression would not only remove the performance degradation we observed, but might well increase the effective dkk transfer rate beyond that obtainable from a system without compression.

We describe a theory of authentication and a system that implements it. Our theory is based on the notion of principal and a “speaks for” relation between principals. A simple principal either has a name or is a communication channel; a compound principal can express an adopted role or delegated authority. The theory shows how to reason about a principal's authority by deducing the other principals that it can speak for; authenticating a channel is one important application. We use the theory to explain many existing and proposed security mechanisms. In particular, we describe the system we have built. It passes principals efficiently as arguments or results of remote procedure calls, and it handles public and shared key encryption, name lookup in a large name space, groups of principals, program loading, delegation, access control, and revocation.

We describe a theory of authentication and a system that implements it. Our theory is based on the notion of principal and a "speaks for" relation between principals. A simple principal either has a name or is a communication channel; a compound principal can express an adopted role or delegation of authority. The theory explains how to reason about a principal's authority by deducing the other principals that it can speak for; authenticating a channel is one important application. We use the theory to explain many existing and proposed mechanisms for security. In particular, we describe the system we have built. It passes principals efficiently as arguments or results of remote procedure calls, and it handles public and shared key encryption, name lookup in a large name space, groups of principals, loading programs, delegation, access control, and revocation.

Autonet is a self-configuring local area network composed of switches interconnected by 100 Mb/s, full-duplex, point-to-point links. The switches contain 12 ports that are internally connected by a full crossbar. Switches use cut-through to achieve a packet forwarding latency as low as 2 ms/switch. Any switch port can be cabled to any other switch port or to a host network controller. A processor in each switch monitors the network's physical configuration. A distributed algorithm running on the switch processor computes the routes packets are to follow and fills in the packet forwarding table in each switch. With Autonet, distinct paths through the set of network links can carry packets in parallel, allowing many pairs of hosts to communicate simultaneously at full link bandwidth. A 30-switch network with more than 100 hosts has been the service network for Digital's Systems Research Center since February 1990.>

Authentication protocols are the basis of security in many distributed systems, and it is therefore essential to ensure that these protocols function correctly. Unfortunately, their design has been extremely error prone. Most of the protocols found in the literature contain redundancies or security flaws. A simple logic has allowed us to describe the beliefs of trustworthy parties involved in authentication protocols and the evolution of these beliefs as a consequence of communication. We have been able to explain a variety of authentication protocols formally, to discover subtleties and errors in them, and to suggest improvements. In this paper we present the logic and then give the results of our analysis of four published protocols, chosen either because of their practical importance or because they serve to illustrate our method.

In this paper we report on the performance of the remote procedure call (RPC) implementation for the Firefly multiprocessor and analyze the implementation to account precisely for all measured latency. From the analysis and measurements, we estimate how much faster RPC could be if certain improvements were made. The elapsed time for an intermachine call to a remote procedure that accepts no arguments and produces no results is 2.66 ms. The elapsed time for an RPC that has a single 1440-byte result (the maximum result that will fit in a single packet) is 6.35 ms. Maximum intermachine throughput of application program data using RPC is 4.65 Mbits/s, achieved with four threads making parallel RPCs that return the maximum-size result that fits in a single RPC result packet. CPU utilization at maximum throughput is about 1.2 CPU seconds per second on the calling machine and a little less on the server. These measurements are for RPCs from user space on one machine to user space on another, using the installed system and a 10 Mbit/s Ethernet. The RPC packet exchange protocol is built on IP/UDP, and the times include calculating and verifying UDP checksums. The Fireflies used in the tests had 5 MicroVAX II processors and a DEQNA Ethernet controller.

Authentication protocols are the basis of security in many distributed systems, and it is therefore essential to ensure that these protocols function correctly. Unfortunately, their design has been extremely error prone. Most of the protocols found in the literature contain redundancies or security flaws.

In this paper, we report on the performance of the remote procedure call implementation for the Firefly multiprocessor and analyze the implementation to account precisely for all measured latency. From the analysis and measurements, we estimate how much faster RPC could be if certain improvements were made.