Section: New Results

High-performance computing on next generation architectures

Memory optimization for the training phase of deep convolutional networks

Training Deep Neural Networks is known to be an expensive operation, both in terms of computational cost and memory load. Indeed, during training, all intermediate layer outputs (called activations) computed during the forward phase must be stored until the corresponding gradient has been computedin the backward phase. These memory requirements sometimes prevent to consider larger batch sizes and deeper networks, so that they can limit both convergence speed and accuracy. Recent works have proposed to offload some of the computed forward activations from the memory of the GPU to the memory of the CPU. This requires to determine which activations should be offloaded and when these transfers from and to the memory of the GPU should take place. In [28], We prove that this problem is NP-hard in the strong sense, and we propose two heuristics based on relaxations of the problem. We perform extensive experimental evaluation on standard Deep Neural Networks. We compare the performance of our heuristics against previous approaches from the literature, showing that they achieve much better performance in a wide variety of situations.

In [23], we also introduce a new activation checkpointing method which allows to significantly decrease memory usage when training Deep Neural Networks with the back-propagation algorithm. Similarly to checkpointing techniques coming from the literature on Automatic Differentiation, it consists in dynamically selecting the forward activations that are saved during the training phase, and then automatically recomputing missing activations from those previously recorded. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs, or recording the complete history of operations that produced the outputs (this uses more memory, but requires fewer recomputations in the backward phase), and we provide an algorithm to compute the optimal computation sequence for this model. This paper also describes a PyTorch implementation that processes the entire chain, dealing with any sequential DNN whose internal layers may be arbitrarily complex and automatically executing it according to the optimal checkpointing strategy computed given a memory limit. Through extensive experiments, we show that our implementation consistently outperforms existing checkpointing approaches for a large class of networks, image sizes and batch sizes.

In [4], [24], we consider the problem of optimally scheduling the backpropagation of Deep Join Networks. Deep Learning training memory needs can prevent the user to consider large models and large batch sizes. In this work, we propose to use techniques from memory-aware scheduling and Automatic Differentiation (AD) to execute a backpropagation graph with a bounded memory requirement at the cost of extra recomputations. The case of a single homogeneous chain, i.e. the case of a network whose all stages are identical and form a chain, is well understood and optimal solutions have been proposed in the AD literature. The networks encountered in practice in the context of Deep Learning are much more diverse, both in terms of shape and heterogeneity. In this work, we define the class of backpropagation graphs, and extend those on which one can compute in polynomial time a solution that minimizes the total number of recomputations. In particular we consider join graphs which correspond to models such as Siamese or Cross Modal Networks.

Sizing and Partitioning Strategies for Burst-Buffers to Reduce IO Contention

Burst-Buffers are high throughput and small size storage which are being used as an intermediate storage between the PFS (Parallel File System) and the computational nodes of modern HPC systems. They can allow to hinder to contention to the PFS, a shared resource whose read and write performance increase slower than processing power in HPC systems. A second usage is to accelerate data transfers and to hide the latency to the PFS. In this work, we concentrate on the first usage. We propose a model for Burst-Buffers and application transfers. We consider the problem of dimensioning and sharing the Burst-Buffers between several applications. This dimensioning can be done either dynamically or statically. The dynamic allocation considers that any application can use any available portion of the Burst-Buffers. The static allocation considers that when a new application enters the system, it is assigned some portion of the Burst-Buffers, which cannot be used by the other applications until that application leaves the system and its data is purged from it. We show that the general sharing problem to guarantee fair performance for all applications is an NP-Complete problem. We propose a polynomial time algorithms for the special case of finding the optimal buffer size such that no application is slowed down due to PFS contention, both in the static and dynamic cases. Finally, we provide evaluations of our algorithms in realistic settings. We use those to discuss how to minimize the overhead of the static allocation of buffers compared to the dynamic allocation. More information on these results can be found in [9].

Efficient Ordering of Kernel Submission on GPUs

In distributed memory systems, it is paramount to develop strategies to overlap the data transfers between memory nodes with the computations in order to exploit their full potential. In [11], we consider the problem of determining the order of data transfers between two memory nodes for a set of independent tasks with the objective of minimizing the makespan. We prove that, with limited memory capacity, the problem of obtaining the optimal data transfer order is NP-complete. We propose several heuristics to determine this order and discuss the conditions that might be favorable to different heuristics. We analyze our heuristics on traces obtained by running two molecular chemistry kernels, namely, Hartree–Fock (HF) and Coupled Cluster Singles Doubles (CCSD), on 10 nodes of an HPC system. Our results show that some of our heuristics achieve significant overlap for moderate memory capacities and resulting in makespans that are very close to the lower bound.

Concurrent kernel execution is a relatively new feature in modern GPUs, which was designed to improve hardware utilization and the overall system throughput. However, the decision on the simultaneous execution of tasks is performed by the hardware with a leftover policy, that assigns as many resources as possible for one task and then assigns the remaining resources to the next task. This can lead to unreasonable use of resources. In [30], we tackle the problem of co-scheduling for GPUs with and without preemption, with the focus on determining the kernels submission order to reduce the number of preemptions and the kernels makespan, respectively. We propose a graph-based theoretical model to build preemptive and non-preemptive schedules. We show that the optimal preemptive makespan can be computed by solving a Linear Program in polynomial time, and we propose an algorithm based on this solution which minimizes the number of preemptions. We also propose an algorithm that transforms a preemptive solution of optimal makespan into a non-preemptive solution with the smallest possible preemption overhead. We show, however, that finding the minimal amount of preemptions among all preemptive solutions of optimal makespan is a NP-hard problem, and computing the optimal non-preemptive schedule is also NP-hard. In addition, we study the non-preemptive problem, without searching first for a good preemptive solution, and present a Mixed Integer Linear Program solution to this problem. We performed experiments on real-world GPU applications and our approach can achieve optimal makespan by preempting 6 to 9% of the tasks. Our non-preemptive approach, on the other side, obtains makespan within 2.5% of the optimal preemptive schedules, while previous approaches exceed the preemptive makespan by 5 to 12%.

Scheduling Tasks on Two Types of Resources

We consider the problem of scheduling task graphs on two types of unrelated resources, which arises in the context of task-based runtime systems on modern platforms containing CPUs and GPUs. In [10], we focus on an algorithm named HeteroPrio, which was originally introduced as an efficient heuristic for a particular application. HeteroPrio is an adaptation of the well known list scheduling algorithm, in which the tasks are picked by the resources in the order of their acceleration factor. This algorithm is augmented with a spoliation mechanism: a task assigned by the list algorithm can later on be reassigned to a different resource if it allows to finish this task earlier. We propose here the first theoretical analysis of the HeteroPrio algorithm in the presence of dependencies. More specifically, if the platform contains m and n processors of each type, we show that the worst-case approximation ratio of HeteroPrio is between $1+max(m/n,n/m)$ and $2+max(m/n,n/m)$. Our proof structure allows to precisely identify the necessary conditions on the spoliation strategy to obtain such a guarantee. We also present an in-depth experimental analysis, comparing several such spoliation strategies, and comparing HeteroPrio with other algorithms from the literature. Although the worst case analysis shows the possibility of pathological behavior, HeteroPrio is able to produce, in very reasonable time, schedules of significantly better quality.

The evolution in the design of modern parallel platforms leads to revisit the scheduling jobs on distributed heterogeneous resources. We contribute to [31], a survey whose goal is to present the main existing algorithms, to classify them based on their underlying principles and to propose unified implementations to enable their fair comparison, both in terms of running time and quality of schedules, on a large set of common benchmarks that we made available for the community [27]. Beyond this comparison, our goal is also to understand the main difficulties that heterogeneity conveys and the shared principles that guide the design of efficient algorithms.

Data-Locality Aware Tasks Scheduling with Replicated Inputs

In [5], we consider the influence on data-locality of the replication of data files, as automatically performed by Distributed File Systems such as HDFS. Replication is known to have a crucial impact on data locality in addition to system fault tolerance. Indeed, intuitively, having more replicas of the same input file gives more opportunities for this task to be processed locally, i.e. without any input file transfer. Given the practical importance of this problem, a vast literature has been proposed to schedule tasks, based on a random placement of replicated input files. Our goal in this paper is to study the performance of these algorithms, both in terms of makespan minimization (minimize the completion time of the last task when non-local processing is forbidden) and communication minimization (minimize the number of non-local tasks when no idle time on resources is allowed). In the case of homogenous tasks, we are able to prove, using models based on "balls into bins" and "power of two choices" problems, that the well known good behavior of classical strategies can be theoretically grounded. Going further, we even establish that it is possible, using semi-matchings theory, to find the optimal solution in very small time. We also use known graph-orientation results to prove that this optimal solution is indeed near-perfect with strong probability. In the more general case of heterogeneous tasks, we propose heuristics solutions both in the clairvoyant and non-clairvoyant cases (i.e. task length is known in advance or not), and we evaluate them through simulations, using actual traces of a Hadoop cluster.