arXiv subjects

Marco Canini

Publications and source records attributed to Marco Canini.

At least 55 records · Page 3Linked to original sources

arXiv2021

RDMA is Turing complete, we just did not know it yet!

It is becoming increasingly popular for distributed systems to exploit offload to reduce load on the CPU. Remote Direct Memory Access (RDMA) offload, in particular, has become popular. However, RDMA still requires CPU intervention for complex offloads that go beyond simple remote memory access. As such, the offload potential is limited and RDMA-based systems usually have to work around such limitations. We present RedN, a principled, practical approach to implementing complex RDMA offloads, without requiring any hardware modifications. Using self-modifying RDMA chains, we lift the existing RDMA verbs interface to a Turing complete set of programming abstractions. We explore what is possible in terms of offload complexity and performance with a commodity RDMA NIC. We show how to integrate these RDMA chains into applications, such as the Memcached key-value store, allowing us to offload complex tasks such as key lookups. RedN can reduce the latency of key-value get operations by up to 2.6x compared to state-of-the-art KV designs that use one-sided RDMA primitives (e.g., FaRM-KV), as well as traditional RPC-over-RDMA approaches. Moreover, compared to these baselines, RedN provides performance isolation and, in the presence of contention, can reduce latency by up to 35x while providing applications with failure resiliency to OS and process crashes.

cs.DC↗

arXiv2021

Rethinking gradient sparsification as total error minimization

Gradient compression is a widely-established remedy to tackle the communication bottleneck in distributed training of large deep neural networks (DNNs). Under the error-feedback framework, Top-$k$ sparsification, sometimes with $k$ as little as $0.1\%$ of the gradient size, enables training to the same model quality as the uncompressed case for a similar iteration count. From the optimization perspective, we find that Top-$k$ is the communication-optimal sparsifier given a per-iteration $k$ element budget. We argue that to further the benefits of gradient sparsification, especially for DNNs, a different perspective is necessary -- one that moves from per-iteration optimality to consider optimality for the entire training. We identify that the total error -- the sum of the compression errors for all iterations -- encapsulates sparsification throughout training. Then, we propose a communication complexity model that minimizes the total error under a communication budget for the entire training. We find that the hard-threshold sparsifier, a variant of the Top-$k$ sparsifier with $k$ determined by a constant hard-threshold, is the optimal sparsifier for this model. Motivated by this, we provide convex and non-convex convergence analyses for the hard-threshold sparsifier with error-feedback. Unlike with Top-$k$ sparsifier, we show that hard-threshold has the same asymptotic convergence and linear speedup property as SGD in the convex case and has no impact on the data-heterogeneity in the non-convex case. Our diverse experiments on various DNNs and a logistic regression model demonstrated that the hard-threshold sparsifier is more communication-efficient than Top-$k$.

cs.LG↗

arXiv2021

With Great Freedom Comes Great Opportunity: Rethinking Resource Allocation for Serverless Functions

Current serverless offerings give users a limited degree of flexibility for configuring the resources allocated to their function invocations by either coupling memory and CPU resources together or providing no knobs at all. These configuration choices simplify resource allocation decisions on behalf of users, but at the same time, create deployments that are resource inefficient. In this paper, we take a principled approach to the problem of resource allocation for serverless functions, allowing this choice to be made in an automatic way that leads to the best combination of performance and cost. In particular, we systematically explore the opportunities that come with decoupling memory and CPU resource allocations and also enabling the use of different VM types. We find a rich trade-off space between performance and cost. The provider can use this in a number of ways: from exposing all these parameters to the user, to eliciting preferences for performance and cost from users, or by simply offering the same performance with lower cost. This flexibility can also enable the provider to optimize its resource utilization and enable a cost-effective service with predictable performance. Our results show that, by decoupling memory and CPU allocation, there is potential to have up to 40% lower execution cost than the preset coupled configurations that are the norm in current serverless offerings. Similarly, making the correct choice of VM instance type can provide up to 50% better execution time. Furthermore, we demonstrate that providers can utilize different instance types for the same functions to maximize resource utilization while providing performance within 10-20% of the best resource configuration for each respective function.

cs.DC↗

arXiv2021

AutoLRS: Automatic Learning-Rate Schedule by Bayesian Optimization on the Fly

The learning rate (LR) schedule is one of the most important hyper-parameters needing careful tuning in training DNNs. However, it is also one of the least automated parts of machine learning systems and usually costs significant manual effort and computing. Though there are pre-defined LR schedules and optimizers with adaptive LR, they introduce new hyperparameters that need to be tuned separately for different tasks/datasets. In this paper, we consider the question: Can we automatically tune the LR over the course of training without human involvement? We propose an efficient method, AutoLRS, which automatically optimizes the LR for each training stage by modeling training dynamics. AutoLRS aims to find an LR applied to every $τ$ steps that minimizes the resulted validation loss. We solve this black-box optimization on the fly by Bayesian optimization (BO). However, collecting training instances for BO requires a system to evaluate each LR queried by BO's acquisition function for $τ$ steps, which is prohibitively expensive in practice. Instead, we apply each candidate LR for only $τ'\llτ$ steps and train an exponential model to predict the validation loss after $τ$ steps. This mutual-training process between BO and the loss-prediction model allows us to limit the training steps invested in the BO search. We demonstrate the advantages and the generality of AutoLRS through extensive experiments of training DNNs for tasks from diverse domains using different optimizers. The LR schedules auto-generated by AutoLRS lead to a speedup of $1.22\times$, $1.43\times$, and $1.5\times$ when training ResNet-50, Transformer, and BERT, respectively, compared to the LR schedules in their original papers, and an average speedup of $1.31\times$ over state-of-the-art heavily-tuned LR schedules.

cs.LG↗

arXiv2021

An Efficient Statistical-based Gradient Compression Technique for Distributed Training Systems

The recent many-fold increase in the size of deep neural networks makes efficient distributed training challenging. Many proposals exploit the compressibility of the gradients and propose lossy compression techniques to speed up the communication stage of distributed training. Nevertheless, compression comes at the cost of reduced model quality and extra computation overhead. In this work, we design an efficient compressor with minimal overhead. Noting the sparsity of the gradients, we propose to model the gradients as random variables distributed according to some sparsity-inducing distributions (SIDs). We empirically validate our assumption by studying the statistical characteristics of the evolution of gradient vectors over the training process. We then propose Sparsity-Inducing Distribution-based Compression (SIDCo), a threshold-based sparsification scheme that enjoys similar threshold estimation quality to deep gradient compression (DGC) while being faster by imposing lower compression overhead. Our extensive evaluation of popular machine learning benchmarks involving both recurrent neural network (RNN) and convolution neural network (CNN) models shows that SIDCo speeds up training by up to 41:7%, 7:6%, and 1:9% compared to the no-compression baseline, Topk, and DGC compressors, respectively.

cs.LG↗

arXiv2021

On the Impact of Device and Behavioral Heterogeneity in Federated Learning

Federated learning (FL) is becoming a popular paradigm for collaborative learning over distributed, private datasets owned by non-trusting entities. FL has seen successful deployment in production environments, and it has been adopted in services such as virtual keyboards, auto-completion, item recommendation, and several IoT applications. However, FL comes with the challenge of performing training over largely heterogeneous datasets, devices, and networks that are out of the control of the centralized FL server. Motivated by this inherent setting, we make a first step towards characterizing the impact of device and behavioral heterogeneity on the trained model. We conduct an extensive empirical study spanning close to 1.5K unique configurations on five popular FL benchmarks. Our analysis shows that these sources of heterogeneity have a major impact on both model performance and fairness, thus sheds light on the importance of considering heterogeneity in FL system design.

cs.LG↗

arXiv2020

Scaling Distributed Machine Learning with In-Network Aggregation

Training machine learning models in parallel is an increasingly important workload. We accelerate distributed parallel training by designing a communication primitive that uses a programmable switch dataplane to execute a key step of the training process. Our approach, SwitchML, reduces the volume of exchanged data by aggregating the model updates from multiple workers in the network. We co-design the switch processing with the end-host protocols and ML frameworks to provide an efficient solution that speeds up training by up to 5.5$\times$ for a number of real-world benchmark models.

cs.DC↗

arXiv2020

Assise: Performance and Availability via NVM Colocation in a Distributed File System

The adoption of very low latency persistent memory modules (PMMs) upends the long-established model of disaggregated file system access. Instead, by colocating computation and PMM storage, we can provide applications much higher I/O performance, sub-second application failover, and strong consistency. To demonstrate this, we built the Assise distributed file system, based on a persistent, replicated coherence protocol for managing a set of server-colocated PMMs as a fast, crash-recoverable cache between applications and slower disaggregated storage, such as SSDs. Unlike disaggregated file systems, Assise maximizes locality for all file IO by carrying out IO on colocated PMM whenever possible and minimizes coherence overhead by maintaining consistency at IO operation granularity, rather than at fixed block sizes. We compare Assise to Ceph/Bluestore, NFS, and Octopus on a cluster with Intel Optane DC PMMs and SSDs for common cloud applications and benchmarks, such as LevelDB, Postfix, and FileBench. We find that Assise improves write latency up to 22x, throughput up to 56x, fail-over time up to 103x, and scales up to 6x better than its counterparts, while providing stronger consistency semantics. Assise promises to beat the MinuteSort world record by 1.5x.

cs.DC↗

arXiv2020

Consistent SDNs through Network State Fuzzing

The conventional wisdom is that a software-defined network (SDN) operates under the premise that the logically centralized control plane has an accurate representation of the actual data plane state. Unfortunately, bugs, misconfigurations, faults or attacks can introduce inconsistencies that undermine correct operation. Previous work in this area, however, lacks a holistic methodology to tackle this problem and thus, addresses only certain parts of the problem. Yet, the consistency of the overall system is only as good as its least consistent part. Motivated by an analogy of network consistency checking with program testing, we propose to add active probe-based network state fuzzing to our consistency check repertoire. Hereby, our system, PAZZ, combines production traffic with active probes to periodically test if the actual forwarding path and decision elements (on the data plane) correspond to the expected ones (on the control plane). Our insight is that active traffic covers the inconsistency cases beyond the ones identified by passive traffic. PAZZ prototype was built and evaluated on topologies of varying scale and complexity. Our results show that PAZZ requires minimal network resources to detect persistent data plane faults through fuzzing and localize them quickly while outperforming baseline approaches.

cs.NI↗

arXiv2019

On the Discrepancy between the Theoretical Analysis and Practical Implementations of Compressed Communication for Distributed Deep Learning

Compressed communication, in the form of sparsification or quantization of stochastic gradients, is employed to reduce communication costs in distributed data-parallel training of deep neural networks. However, there exists a discrepancy between theory and practice: while theoretical analysis of most existing compression methods assumes compression is applied to the gradients of the entire model, many practical implementations operate individually on the gradients of each layer of the model. In this paper, we prove that layer-wise compression is, in theory, better, because the convergence rate is upper bounded by that of entire-model compression for a wide range of biased and unbiased compression methods. However, despite the theoretical bound, our experimental study of six well-known methods shows that convergence, in practice, may or may not be better, depending on the actual trained model and compression ratio. Our findings suggest that it would be advantageous for deep learning frameworks to include support for both layer-wise and entire-model compression.

cs.DC↗

arXiv2019

Direct Nonlinear Acceleration

Optimization acceleration techniques such as momentum play a key role in state-of-the-art machine learning algorithms. Recently, generic vector sequence extrapolation techniques, such as regularized nonlinear acceleration (RNA) of Scieur et al., were proposed and shown to accelerate fixed point iterations. In contrast to RNA which computes extrapolation coefficients by (approximately) setting the gradient of the objective function to zero at the extrapolated point, we propose a more direct approach, which we call direct nonlinear acceleration (DNA). In DNA, we aim to minimize (an approximation of) the function value at the extrapolated point instead. We adopt a regularized approach with regularizers designed to prevent the model from entering a region in which the functional approximation is less precise. While the computational cost of DNA is comparable to that of RNA, our direct approach significantly outperforms RNA on both synthetic and real-world datasets. While the focus of this paper is on convex problems, we obtain very encouraging results in accelerating the training of neural networks.

math.OC↗

arXiv2019

Partitioned Paxos via the Network Data Plane

Consensus protocols are the foundation for building fault-tolerant, distributed systems, and services. They are also widely acknowledged as performance bottlenecks. Several recent systems have proposed accelerating these protocols using the network data plane. But, while network-accelerated consensus shows great promise, current systems suffer from an important limitation: they assume that the network hardware also accelerates the application itself. Consequently, they provide a specialized replicated service, rather than providing a general-purpose high-performance consensus that fits any off-the-shelf application. To address this problem, this paper proposes Partitioned Paxos, a novel approach to network-accelerated consensus. The key insight behind Partitioned Paxos is to separate the two aspects of Paxos, agreement, and execution, and optimize them separately. First, Partitioned Paxos uses the network forwarding plane to accelerate agreement. Then, it uses state partitioning and parallelization to accelerate execution at the replicas. Our experiments show that using this combination of data plane acceleration and parallelization, Partitioned Paxos is able to provide at least x3 latency improvement and x11 throughput improvement for a replicated instance of a RocksDB key-value store.

cs.DC↗

arXiv2019

Measurements As First-class Artifacts

The emergence of programmable switches has sparked a significant amount of work on new techniques to perform more powerful measurement tasks, for instance, to obtain fine-grained traffic and performance statistics. Previous work has focused on the efficiency of these measurements alone and has neglected flexibility, resulting in solutions that are hard to reuse or repurpose and that often overlap in functionality or goals. In this paper, we propose the use of a set of reusable primitive building blocks that can be composed to express measurement tasks in a concise and simple way. We describe the rationale for the design of our primitives, that we have named MAFIA (Measurements As FIrst-class Artifacts), and using several examples we illustrate how they can be combined to realize a comprehensive range of network measurement tasks. Writing MAFIA code does not require expert knowledge of low-level switch architecture details. Using a prototype implementation of MAFIA, we demonstrate the applicability of our approach and show that the use of our primitives results in compiled code that is comparable in size and resource usage with manually written specialized P4 code and can be run in current hardware.

cs.NI↗

arXiv2018

Prelude: Ensuring Inter-Domain Loop-Freedom in~SDN-Enabled Networks

Software-Defined-eXchanges (SDXes) promise to tackle the timely quest of bringing improving the inter-domain routing ecosystem through SDN deployment. Yet, the naive deployment of SDN on the Internet raises concerns about the correctness of the inter-domain data-plane. By allowing operators to deflect traffic from the default BGP route, SDN policies are susceptible of creating permanent forwarding loops invisible to the control-plane. In this paper, we propose a system, called Prelude, for detecting SDN-induced forwarding loops between SDXes with high accuracy without leaking the private routing information of network operators. To achieve this, we leverage Secure Multi-Party Computation (SMPC) techniques to build a novel and general privacy-preserving primitive that detects whether any subset of SDN rules might affect the same portion of traffic without learning anything about those rules. We then leverage that primitive as the main building block of a distributed system tailored to detect forwarding loops among any set of SDXes. We leverage the particular nature of SDXes to further improve the efficiency of our SMPC solution. The number of valid SDN rules, i.e., not creating loops, rejected by our solution is 100x lower than previous privacy-preserving solutions, and also provides better privacy guarantees. Furthermore, our solution naturally provides network operators with some hindsight on the cost of the deflected paths.

cs.NI↗

arXiv2017

Sonata: Query-Driven Network Telemetry

Operating networks depends on collecting and analyzing measurement data. Current technologies do not make it easy to do so, typically because they separate data collection (e.g., packet capture or flow monitoring) from analysis, producing either too much data to answer a general question or too little data to answer a detailed question. In this paper, we present Sonata, a network telemetry system that uses a uniform query interface to drive the joint collection and analysis of network traffic. Sonata takes the advantage of two emerging technologies---streaming analytics platforms and programmable network devices---to facilitate joint collection and analysis. Sonata allows operators to more directly express network traffic analysis tasks in terms of a high-level language. The underlying runtime partitions each query into a portion that runs on the switch and another that runs on the streaming analytics platform iteratively refines the query to efficiently capture only the traffic that pertains to the operator's query, and exploits sketches to reduce state in switches in exchange for more approximate results. Through an evaluation of a prototype implementation, we demonstrate that Sonata can support a wide range of network telemetry tasks with less state in the network, and lower data rates to streaming analytics systems, than current approaches can achieve.

cs.NI↗

arXiv2017

Decentralized Consistent Network Updates in SDN with ez-Segway

We present ez-Segway, a decentralized mechanism to consistently and quickly update the network state while preventing forwarding anomalies (loops and black-holes) and avoiding link congestion. In our design, the centralized SDN controller only pre-computes information needed by the switches during the update execution. This information is distributed to the switches, which use partial knowledge and direct message passing to efficiently realize the update. This separation of concerns has the key benefit of improving update performance as the communication and computation bottlenecks at the controller are removed. Our evaluations via network emulations and large-scale simulations demonstrate the efficiency of ez-Segway, which compared to a centralized approach, improves network update times by up to 45% and 57% at the median and the 99th percentile, respectively. A deployment of a system prototype in a real OpenFlow switch and an implementation in P4 demonstrate the feasibility and low overhead of implementing simple network update functionality within switches.

cs.NI↗

arXiv2016

Network Hardware-Accelerated Consensus

Consensus protocols are the foundation for building many fault-tolerant distributed systems and services. This paper posits that there are significant performance benefits to be gained by offering consensus as a network service (CAANS). CAANS leverages recent advances in commodity networking hardware design and programmability to implement consensus protocol logic in network devices. CAANS provides a complete Paxos protocol, is a drop-in replacement for software-based implementations of Paxos, makes no restrictions on network topologies, and is implemented in a higher-level, data-plane programming language, allowing for portability across a range of target devices. At the same time, CAANS significantly increases throughput and reduces latency for consensus operations. Consensus logic executing in hardware can transmit consensus messages at line speed, with latency only slightly higher than simply forwarding packets.

cs.DC↗

arXiv2015

Paxos Made Switch-y

This paper describes an implementation of the well-known consensus protocol, Paxos, in the P4 programming language. P4 is a language for programming the behavior of network forwarding devices (i.e., the network data plane). Moving consensus logic into network devices could significantly improve the performance of the core infrastructure and services in data centers. Moreover, implementing Paxos in P4 provides a critical use case and set of requirements for data plane language designers. In the long term, we imagine that consensus could someday be offered as a network service, just as point-to-point communication is provided today.

cs.DC↗