Identifying performance bottlenecks in a parallel application is tedious, especially because it requires analyzing the behaviour of various software components, as bottlenecks may have several causes and symptoms. For example, a load imbalance may cause long MPI waiting times, or contention on disk may degrade the performance of I/O operations. Detecting a performance problem means investigating the execution of an application and applying several performance analysis techniques. To do so, one can use a tracing tool to collect information describing the behaviour of the application. At the end of the execution, a trace file in a specific format is available to the application user, which can be used to conduct a complete post-mortem investigation. Several challenges emerge from the generation and use of traces. Tracing applications may alter the performance of the application, and can create thousands of heavy trace files, especially at a large scale. Most importantly, the post-mortem analysis needs to load these thousands of trace files in memory, and process them. This quickly becomes impractical for large scale applications, as memory gets exhausted and the number of opened files exceeds the system capacity. In this paper, we propose PALLAS, a generic trace format tailored for conducting various post-mortem performance analysis of traces describing large executions of HPC applications. During the execution of the application, PALLAS collects events and detects their repetitions on-the-fly. When storing the trace to disk, PALLAS groups the data from similar events or groups of events together in order to later speed up trace reading. We demonstrate that the PALLAS online detection of the program structure does not significantly degrade the performance of the applications. Moreover, the PALLAS format allows faster trace analysis compared to other evaluated trace formats. Overall, the PALLAS trace format allows an interactive analysis of a trace that is required when a user investigates a performance problem.

P4ce is the first replication protocol that exhibits the same latency and requires the same network capacity as sending data to a single server. P4ce builds upon previous RDMA-based consensus protocols. They achieve consensus with a single network round-trip, but with a reduced network throughput. P4ce also achieves consensus with a single round-trip, but without degrading throughput by decoupling the consensus decisions from the RDMA communications. The decision part of the consensus protocol runs on a commodity server, but the communication part of P4ce is fully implemented on a programmable switch, which replicates data and aggregates the acknowledgements in the network, avoiding the throughput bottleneck at the leader. Although simple in its principle, the implementation of P4ce raises many challenging issues, notably caused by the complexity of RDMA and the underlying network protocols, the intricacies of packet rewriting during replication and aggregation, and the restricted set of operations that can be implemented at wire speed in the programmable switch. We implemented P4ce and deployed it on a commercially-available Intel Tofino switch, achieving up to 4x better through-put and better latency than state-of-the-art consensus protocols.

Partitioning a multi-threaded application between a secure and a non-secure memory zone remains a challenge. The current tools rely on data flow analysis techniques, which are unable to handle multi-threaded C or C++ applications. To avoid this limitation, we propose to trade the ease-of-use of data flow analysis for another language construct: explicit secure typing. With secure typing, as with data flow analysis, the developer annotates memory locations that contain sensitive values. However, instead of analyzing how the sensitive values flow, we propose to use these annotations to only check typing rules, such as ensuring that the code never stores a sensitive value in an unsafe memory location. By avoiding data flow analysis, the developer has to annotate more memory locations, but the partitioning tool can handle multi-threaded C and C++ applications.

Trusted execution environments like Intel SGX provide enclaves, which offer strong security guarantees for applications. Running entire applications inside enclaves is possible, but this approach leads to a large trusted computing base (TCB). As such, various tools have been developed to partition programs written in languages such as C or Java into trusted and untrusted parts, which are run in and out of enclaves respectively. However, those tools depend on language-specific taint-analysis and partitioning techniques. They cannot be reused for other languages and there is thus a need for tools that transcend this language barrier.

This paper introduces NVCACHE, an approach that uses a non-volatile main memory (NVMM) as a write cache to improve the write performance of legacy applications. We compare NVCACHE against file systems tailored for NVMM (Ext4-DAX and NOVA) and with I/O-heavy applications (SQLite, RocksDB). Our evaluation shows that NVCACHE reaches the performance level of the existing state-of-the-art systems for NVMM, but without their limitations: NVCACHE does not limit the size of the stored data to the size of the NVMM, and works transparently with unmodified legacy applications, providing additional persistence guarantees even when their source code is not available.

The popularity of the Java programming language has led to its wide adoption in cloud computing infrastructures. However, Java applications running in untrusted clouds are vulnerable to various forms of privileged attacks. The emergence of trusted execution environments (TEEs) such as Intel SGX mitigates this problem. TEEs protect code and data in secure enclaves inaccessible to untrusted software, including the kernel and hypervisors. To efficiently use TEEs, developers must manually partition their applications into trusted and untrusted parts, in order to reduce the size of the trusted computing base (TCB) and minimise the risks of security vulnerabilities. However, partitioning applications poses two important challenges: (i) ensuring efficient object communication between the partitioned components, and (ii) ensuring the consistency of garbage collection between the parts, especially with memory-managed languages such as Java. We present Montsalvat, a tool which provides a practical and intuitive annotation-based partitioning approach for Java applications destined for secure enclaves. Montsalvat provides an RMI-like mechanism to ensure inter-object communication, as well as consistent garbage collection across the partitioned components. We implement Montsalvat with GraalVM native-image, a tool for compiling Java applications ahead-of-time into standalone native executables that do not require a JVM at runtime. Our extensive evaluation with micro- and macro-benchmarks shows our partitioning approach to boost performance in real-world applications up to 6.6x (PalDB) and 2.2x (GraphChi) as compared to solutions that naively include the entire applications in the enclave.

This paper presents J-NVM, a framework to access efficiently Non-Volatile Main Memory (NVMM) in Java. J-NVM offers a fully-fledged interface to persist plain Java objects using failure-atomic blocks. This interface relies internally on proxy objects that intermediate direct off-heap access to NVMM. The framework also provides a library of highly-optimized persistent data types that resist reboots and power failures. We evaluate J-NVM by implementing a persistent backend for the Infinispan data store. Our experimental results, obtained with a TPC-B like benchmark and YCSB, show that J-NVM is consistently faster than other approaches at accessing NVMM in Java.

This paper addresses the problem of efficiently virtualizing NUMA architectures. The major challenge comes from the fact that the hypervisor regularly reconfigures the placement of a virtual machine (VM) over the NUMA topology. However, neither guest operating systems (OSes) nor system runtime libraries (e.g., Hotspot) are designed to consider NUMA topology changes at runtime, leading end user applications to unpredictable performance. This paper presents eXtended Para-Virtualization (XPV), a new principle to efficiently virtualize a NUMA architecture. XPV consists in revisiting the interface between the hypervisor and the guest OS, and between the guest OS and system runtime libraries (SRL) so that they can dynamically take into account NUMA topology changes. The paper presents a methodology for systematically adapting legacy hypervisors, OSes, and SRLs. We have applied our approach with less than 2k line of codes in two legacy hypervisors (Xen and KVM), two legacy guest OSes (Linux and FreeBSD), and three legacy SRLs (Hotspot, TCMalloc, and jemalloc). The evaluation results showed that XPV outperforms all existing solutions by up to 304%.

Understanding the performance of a multi-threaded application is difficult. The threads interfere when they access the same shared resource, which slows down their execution. Unfortunately, current profiling tools report the hardware components or the synchronization primitives that saturate, but they cannot tell if the saturation is the cause of a performance bottleneck. In this paper, we propose a holistic metric able to pinpoint the blocks of code that suffer interference the most, regardless of the interference cause. Our metric uses performance variation as a universal indicator of interference problems. With an evaluation of 27 applications we show that our metric can identify interference problems caused by six different kinds of interference in nine applications. We are able to easily remove seven of the bottlenecks, which leads to a performance improvement of up to nine times.

Novel trusted execution support, as offered by Intel's Software Guard eXtensions (SGX), embeds seamlessly into user space applications by establishing regions of encrypted memory, called enclaves. Enclaves comprise code and data that is executed under special protection of the CPU and can only be accessed via an enclave defined interface. To facilitate the usability of this new system abstraction, Intel offers a software development kit (SGX SDK). While the SDK eases the use of SGX, it misses appropriate programming support for inter-enclave interaction, and demands to hardcode the exact use of trusted execution into applications, which restricts flexibility.

While virtualization only introduces a small overhead on machines with few cores, this is not the case on larger ones. Most of the overhead on the latter machines is caused by the Non-Uniform Memory Access (NUMA) architecture they are using. In order to reduce this overhead, this paper shows how NUMA placement heuristics can be implemented inside Xen. With an evaluation of 29 applications on a 48-core machine, we show that the NUMA placement heuristics can multiply the performance of 9 applications by more than 2.

On contemporary cache-coherent Non-Uniform Memory Access (ccNUMA) architectures, applications with a large memory footprint suffer from the cost of the garbage collector (GC), because, as the GC scans the reference graph, it makes many remote memory accesses, saturating the interconnect between memory nodes. We address this problem with NumaGiC, a GC with a mostly-distributed design. In order to maximise memory access locality during collection, a GC thread avoids accessing a different memory node, instead notifying a remote GC thread with a message; nonetheless, NumaGiC avoids the drawbacks of a pure distributed design, which tends to decrease parallelism. We compare NumaGiC with Parallel Scavenge and NAPS on two different ccNUMA architectures running on the Hotspot Java Virtual Machine of OpenJDK 7. On Spark and Neo4j, two industry-strength analytics applications, with heap sizes ranging from 160GB to 350GB, and on SPECjbb2013 and SPECjbb2005, ourgc improves overall performance by up to 45% over NAPS (up to 94% over Parallel Scavenge), and increases the performance of the collector itself by up to 3.6x over NAPS (up to 5.4x over Parallel Scavenge).

Java class loaders are commonly used in application servers to load, unload and update a set of classes as a unit. However, unloading or updating a class loader can introduce stale references to the objects of the outdated class loader. A stale reference leads to a memory leak and, for an update, to an inconsistency between the outdated classes and their replacements. To detect and eliminate stale references, we propose Incinerator, a Java virtual machine extension that introduces the notion of an outdated class loader. Incinerator detects stale references and sets them to null during a garbage collection cycle. We evaluate Incinerator in the context of the OSGi framework and show that Incinerator correctly detects and eliminates stale references, including a bug in Knopflerfish. We also evaluate the performance of Incinerator with the DaCapo benchmark on VMKit and show that Incinerator has an overhead of at most 3.3%.

Using multiple accelerators, such as GPUs or Xeon Phis, is attractive to improve the performance of large data parallel applications and to increase the size of their workloads. However, writing an application for multiple accelerators remains today challenging because going from a single accelerator to multiple ones indeed requires to deal with potentially non-uniform domain decomposition, inter-accelerator data movements, and dynamic load balancing. Writing such code manually is time consuming and error-prone. In this paper, we propose a new programming tool called STEPOCL along with a new domain specific language designed to simplify the development of an application for multiple accelerators. We evaluate both the performance and the usefulness of STEPOCL with three applications and show that: (i) the performance of an application written with STEPOCL scales linearly with the number of accelerators, (ii) the performance of an application written using STEPOCL competes with a handwritten version, (iii) larger workloads run on multiple devices that do not fit in the memory of a single device, (iv) thanks to STEPOCL, the number of lines of code required to write an application for multiple accelerators is roughly divided by ten.

Today, Java is regularly used to implement large multi-threaded server-class applications that use locks to protect access to shared data. However, understanding the impact of locks on the performance of a system is complex, and thus the use of locks can impede the progress of threads on configurations that were not anticipated by the developer, during specific phases of the execution.

In August 2011, Linux entered its third decade. Ten years before, Chou et al. published a study of faults found by applying a static analyzer to Linux versions 1.0 through 2.4.1. A major result of their work was that the drivers directory contained up to 7 times more of certain kinds of faults than other directories. This result inspired numerous efforts on improving the reliability of driver code. Today, Linux is used in a wider range of environments, provides a wider range of services, and has adopted a new development and release model. What has been the impact of these changes on code quality? To answer this question, we have transported Chou et al.’s experiments to all versions of Linux 2.6; released between 2003 and 2011. We find that Linux has more than doubled in size during this period, but the number of faults per line of code has been decreasing. Moreover, the fault rate of drivers is now below that of other directories, such as arch. These results can guide further development and research efforts for the decade to come. To allow updating these results as Linux evolves, we define our experimental protocol and make our checkers available.

Large-scale multicore architectures create new challenges for garbage collectors (GCs). In particular, throughput-oriented stop-the-world algorithms demonstrate good performance with a small number of cores, but have been shown to degrade badly beyond approximately 8 cores on a 48-core with OpenJDK 7. This negative result raises the question whether the stop-the-world design has intrinsic limitations that would require a radically different approach. Our study suggests that the answer is no, and that there is no compelling scalability reason to discard the existing highly-optimised throughput-oriented GC code on contemporary hardware. This paper studies the default throughput-oriented garbage collector of OpenJDK 7, called Parallel Scavenge. We identify its bottlenecks, and show how to eliminate them using well-established parallel programming techniques. On the SPECjbb2005, SPECjvm2008 and DaCapo 9.12 benchmarks, the improved GC matches the performance of Parallel Scavenge at low core count, but scales well, up to 48~cores.

Omitting resource-release operations in systems error handling code can lead to memory leaks, crashes, and deadlocks. Finding omission faults is challenging due to the difficulty of reproducing system errors, the diversity of system resources, and the lack of appropriate abstractions in the C language. To address these issues, numerous approaches have been proposed that globally scan a code base for common resource-release operations. Such macroscopic approaches are notorious for their many false positives, while also leaving many faults undetected. We propose a novel microscopic approach to finding resource-release omission faults in systems software. Rather than generalizing from the entire source code, our approach focuses on the error-handling code of each function. Using our tool, Hector, we have found over 370 faults in six systems software projects, including Linux, with a 23% false positive rate. Some of these faults allow an unprivileged malicious user to crash the entire system.

In the context of multicore programming, pipeline parallelism is a solution to easily transform a sequential program into a parallel one without requiring a whole rewriting of the code. The OpenMP stream-computing extension presented by Pop and Cohen proposes an extension of OpenMP to handle pipeline parallelism. However, their communication algorithm relies on Multiple-producer-Multiple-Consumer queues, while pipelined applications mostly deal with linear chains of communication, i.e., with only a single producer and a single consumer. To improve the performance of the OpenMP stream-extension, we propose to add a more specialized Single-Producer-Single-Consumer communication algorithm called Batch Queue and to select it for one-to-one communication. Our evaluation shows that Batch Queue is then able to improve the throughput up to a factor 2 on an 8-core machine both for example application and real applications. Our study shows therefore that using specialized and efficient communication algorithms can have a significant impact on the overall performance of pipelined applications.

In 2001, Chou et al. published a study of faults found by applying a static analyzer to Linux versions 1.0 through 2.4.1. A major result of their work was that the drivers directory contained up to 7 times more of certain kinds of faults than other directories. This result inspired a number of development and research efforts on improving the reliability of driver code. Today Linux is used in a much wider range of environments, provides a much wider range of services, and has adopted a new development and release model. What has been the impact of these changes on code quality? Are drivers still a major problem?

Managed Runtime Environments (MRE) are increasingly used for application servers that use large multi-core hardware. We find that the garbage collector is critical for overall performance in this setting. We explore the costs and scalability of the garbage collectors on a contemporary 48-core multiprocessor machine. We present experimental evaluation of the parallel and concurrent garbage collectors present in OpenJDK, a widely-used Java virtual machine. We show that garbage collection represents a substantial amount of an application's execution time, and does not scale well as the number of cores increases. We attempt to identify some critical scalability bottlenecks for garbage collectors.

Massively Multiplayer Online Games (MMOGs) recently emerged as a popular class of applications with millions of users. To offer acceptable gaming experience, such applications need to render the virtual world surrounding the player with a very low latency. However, current state-of-the-art MMOGs based on peer-to-peer overlays fail to satisfy these requirements. This happens because avatar mobility implies many data exchanges through the overlay. As state-of-the-art overlays do not anticipate this mobility, the needed data is not delivered on time, which leads to transient failures at the application level. To solve this problem, we propose Blue Banana, a mechanism that models and predicts avatar movement, allowing the overlay to adapt itself by anticipation to the MMOG needs. Our evaluation is based on large-scale traces derived from Second life. It shows that our anticipation mechanism decreases by 20% the number of transient failures with only a network overhead of 2%.

Mobile ad-hoc networks, MANETs, are self-organized and very dynamic systems where processes have no global knowledge of the system. In this paper, we propose a model that characterizes the dynamics of MANETs in the sense that it considers that paths between nodes are dynamically built and the system can have infinitely many processes but the network may present finite stable partitions. We also propose an algorithm that implements an eventually perfect partition participant detector PD which eventually detects the participant nodes of stable partitions.

Sequential applications can take advantage of multi-core systems by way of pipeline parallelism to improve their performance. In such parallelism, core to core communication overhead is the main limit of speedup. This paper presents BatchQueue, a fast and memory-thrifty core to core communication system based on batch processing of whole cache line. BatchQueue is able to send a 32bit word of data in just 12.5 ns on a Xeon X5472 and only needs 2 full cache lines plus 3 byte-sized variables - each on a different cache line for optimal performance - to work. The characteristics of BatchQueue - high throughput and increased latency resulting from its batch processing - makes it well suited for highly communicative tasks with no real time requirements such as monitoring.

Managed Runtime Environments (MREs), such as the JVM and the CLI, form an attractive environment for program execution, by providing portability and safety, via the use of a bytecode language and automatic memory management, as well as good performance, via just-in-time (JIT) compilation. Nevertheless, developing a fully featured MRE, including e.g. a garbage collector and JIT compiler, is a herculean task. As a result, new languages cannot easily take advantage of the benefits of MREs, and it is difficult to experiment with extensions of existing MRE based languages.

The OSGi framework is a Java-based, centralized, component oriented platform. It is being widely adopted as an execution environment for the development of extensible applications. However, current Java Virtual Machines are unable to isolate components from each other. For instance, a malicious component can freeze the complete platform by allocating too much memory or alter the behavior of other components by modifying shared variables. This paper presents I-JVM, a Java Virtual Machine that provides a lightweight approach to isolation while preserving compatibility with legacy OSGi applications. Our evaluation of I-JVM shows that it solves the 8 known OSGi vulnerabilities that are due to the Java Virtual Machine and that the overhead of I-JVM compared to the JVM on which it is based is below 20%.

Abstract Dynamic flexibility is a major challenge in modern system design to react to context or applicative requirements evolutions. Adapting behaviors may impose substantial code modification across the whole system, in the field, without service interruption and without state loss. This paper presents the JnJVM, a full Java virtual machine (JVM) that satisfies these needs by using dynamic aspect weaving techniques and a component architecture. It supports adding or replacing its own code, while it is running, with no overhead on unmodified code execution. Our measurements reveal similar performance when compared with the monolithic JVM Kaffe. Three illustrative examples show different extension scenarios: (i) modifying the JVMs behavior; (ii) adding capabilities to the JVM; and (iii) modifying applications behavior. Copyright © 2008 John Wiley & Sons, Ltd.