skip to main content


Title: HAL: Computer System for Scalable Deep Learning
We describe the design, deployment and operation of a computer system built to efficiently run deep learning frameworks. The system consists of 16 IBM POWER9 servers with 4 NVIDIA V100 GPUs each, interconnected with Mellanox EDR InfiniBand fabric, and a DDN all-flash storage array. The system is tailored towards efficient execution of the IBM Watson Machine Learning enterprise software stack that combines popular open-source deep learning frameworks. We build a custom management software stack to enable an efficient use of the system by a diverse community of users and provide guides and recipes for running deep learning workloads at scale utilizing all available GPUs. We demonstrate scaling of a PyTorch and TensorFlow based deep neural networks to produce state-of-the-art performance results.  more » « less
Award ID(s):
1725729
NSF-PAR ID:
10190049
Author(s) / Creator(s):
; ; ; ; ; ; ; ; ;
Date Published:
Journal Name:
PEARC '20: Practice and Experience in Advanced Research Computing
Page Range / eLocation ID:
41 to 48
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    The DeepLearningEpilepsyDetectionChallenge: design, implementation, andtestofanewcrowd-sourced AIchallengeecosystem Isabell Kiral*, Subhrajit Roy*, Todd Mummert*, Alan Braz*, Jason Tsay, Jianbin Tang, Umar Asif, Thomas Schaffter, Eren Mehmet, The IBM Epilepsy Consortium◊ , Joseph Picone, Iyad Obeid, Bruno De Assis Marques, Stefan Maetschke, Rania Khalaf†, Michal Rosen-Zvi† , Gustavo Stolovitzky† , Mahtab Mirmomeni† , Stefan Harrer† * These authors contributed equally to this work † Corresponding authors: rkhalaf@us.ibm.com, rosen@il.ibm.com, gustavo@us.ibm.com, mahtabm@au1.ibm.com, sharrer@au.ibm.com ◊ Members of the IBM Epilepsy Consortium are listed in the Acknowledgements section J. Picone and I. Obeid are with Temple University, USA. T. Schaffter is with Sage Bionetworks, USA. E. Mehmet is with the University of Illinois at Urbana-Champaign, USA. All other authors are with IBM Research in USA, Israel and Australia. Introduction This decade has seen an ever-growing number of scientific fields benefitting from the advances in machine learning technology and tooling. More recently, this trend reached the medical domain, with applications reaching from cancer diagnosis [1] to the development of brain-machine-interfaces [2]. While Kaggle has pioneered the crowd-sourcing of machine learning challenges to incentivise data scientists from around the world to advance algorithm and model design, the increasing complexity of problem statements demands of participants to be expert data scientists, deeply knowledgeable in at least one other scientific domain, and competent software engineers with access to large compute resources. People who match this description are few and far between, unfortunately leading to a shrinking pool of possible participants and a loss of experts dedicating their time to solving important problems. Participation is even further restricted in the context of any challenge run on confidential use cases or with sensitive data. Recently, we designed and ran a deep learning challenge to crowd-source the development of an automated labelling system for brain recordings, aiming to advance epilepsy research. A focus of this challenge, run internally in IBM, was the development of a platform that lowers the barrier of entry and therefore mitigates the risk of excluding interested parties from participating. The challenge: enabling wide participation With the goal to run a challenge that mobilises the largest possible pool of participants from IBM (global), we designed a use case around previous work in epileptic seizure prediction [3]. In this “Deep Learning Epilepsy Detection Challenge”, participants were asked to develop an automatic labelling system to reduce the time a clinician would need to diagnose patients with epilepsy. Labelled training and blind validation data for the challenge were generously provided by Temple University Hospital (TUH) [4]. TUH also devised a novel scoring metric for the detection of seizures that was used as basis for algorithm evaluation [5]. In order to provide an experience with a low barrier of entry, we designed a generalisable challenge platform under the following principles: 1. No participant should need to have in-depth knowledge of the specific domain. (i.e. no participant should need to be a neuroscientist or epileptologist.) 2. No participant should need to be an expert data scientist. 3. No participant should need more than basic programming knowledge. (i.e. no participant should need to learn how to process fringe data formats and stream data efficiently.) 4. No participant should need to provide their own computing resources. In addition to the above, our platform should further • guide participants through the entire process from sign-up to model submission, • facilitate collaboration, and • provide instant feedback to the participants through data visualisation and intermediate online leaderboards. The platform The architecture of the platform that was designed and developed is shown in Figure 1. The entire system consists of a number of interacting components. (1) A web portal serves as the entry point to challenge participation, providing challenge information, such as timelines and challenge rules, and scientific background. The portal also facilitated the formation of teams and provided participants with an intermediate leaderboard of submitted results and a final leaderboard at the end of the challenge. (2) IBM Watson Studio [6] is the umbrella term for a number of services offered by IBM. Upon creation of a user account through the web portal, an IBM Watson Studio account was automatically created for each participant that allowed users access to IBM's Data Science Experience (DSX), the analytics engine Watson Machine Learning (WML), and IBM's Cloud Object Storage (COS) [7], all of which will be described in more detail in further sections. (3) The user interface and starter kit were hosted on IBM's Data Science Experience platform (DSX) and formed the main component for designing and testing models during the challenge. DSX allows for real-time collaboration on shared notebooks between team members. A starter kit in the form of a Python notebook, supporting the popular deep learning libraries TensorFLow [8] and PyTorch [9], was provided to all teams to guide them through the challenge process. Upon instantiation, the starter kit loaded necessary python libraries and custom functions for the invisible integration with COS and WML. In dedicated spots in the notebook, participants could write custom pre-processing code, machine learning models, and post-processing algorithms. The starter kit provided instant feedback about participants' custom routines through data visualisations. Using the notebook only, teams were able to run the code on WML, making use of a compute cluster of IBM's resources. The starter kit also enabled submission of the final code to a data storage to which only the challenge team had access. (4) Watson Machine Learning provided access to shared compute resources (GPUs). Code was bundled up automatically in the starter kit and deployed to and run on WML. WML in turn had access to shared storage from which it requested recorded data and to which it stored the participant's code and trained models. (5) IBM's Cloud Object Storage held the data for this challenge. Using the starter kit, participants could investigate their results as well as data samples in order to better design custom algorithms. (6) Utility Functions were loaded into the starter kit at instantiation. This set of functions included code to pre-process data into a more common format, to optimise streaming through the use of the NutsFlow and NutsML libraries [10], and to provide seamless access to the all IBM services used. Not captured in the diagram is the final code evaluation, which was conducted in an automated way as soon as code was submitted though the starter kit, minimising the burden on the challenge organising team. Figure 1: High-level architecture of the challenge platform Measuring success The competitive phase of the "Deep Learning Epilepsy Detection Challenge" ran for 6 months. Twenty-five teams, with a total number of 87 scientists and software engineers from 14 global locations participated. All participants made use of the starter kit we provided and ran algorithms on IBM's infrastructure WML. Seven teams persisted until the end of the challenge and submitted final solutions. The best performing solutions reached seizure detection performances which allow to reduce hundred-fold the time eliptologists need to annotate continuous EEG recordings. Thus, we expect the developed algorithms to aid in the diagnosis of epilepsy by significantly shortening manual labelling time. Detailed results are currently in preparation for publication. Equally important to solving the scientific challenge, however, was to understand whether we managed to encourage participation from non-expert data scientists. Figure 2: Primary occupation as reported by challenge participants Out of the 40 participants for whom we have occupational information, 23 reported Data Science or AI as their main job description, 11 reported being a Software Engineer, and 2 people had expertise in Neuroscience. Figure 2 shows that participants had a variety of specialisations, including some that are in no way related to data science, software engineering, or neuroscience. No participant had deep knowledge and experience in data science, software engineering and neuroscience. Conclusion Given the growing complexity of data science problems and increasing dataset sizes, in order to solve these problems, it is imperative to enable collaboration between people with differences in expertise with a focus on inclusiveness and having a low barrier of entry. We designed, implemented, and tested a challenge platform to address exactly this. Using our platform, we ran a deep-learning challenge for epileptic seizure detection. 87 IBM employees from several business units including but not limited to IBM Research with a variety of skills, including sales and design, participated in this highly technical challenge. 
    more » « less
  2. Graphics processing units (GPUs) manufactured by NVIDIA continue to dominate many fields of research, including real-time GPU-management. NVIDIA’s status as a key enabling technology for deep learning and image processing makes this unsurprising, especially when combined with the company’s push into embedded, safety-critical domains like autonomous driving. NVIDIA’s primary competitor, AMD, has received comparatively little attention, due in part to few embedded offerings and a lack of support from popular deep-learning toolkits. Recently, however, AMD’s ROCm (Radeon Open Compute) software platform was made available to address at least the second of these two issues, but is ROCm worth the attention of safety-critical software developers? In order to answer this question, this paper explores the features and pitfalls of AMD GPUs, focusing on contrasting details with NVIDIA’s GPU hardware and software. We argue that an open software stack such as ROCm may be able to provide much-needed flexibility and reproducibility in the context of real-time GPU research, where new algorithmic or analysis techniques should typically remain agnostic to the underlying GPU architecture. In support of this claim, we summarize how closed-source platforms have obstructed prior research using NVIDIA GPUs, and then demonstrate that AMD may be a viable alternative by modifying components of the ROCm software stack to implement spatial partitioning. Finally, we present a case study using the PyTorch deep-learning framework that demonstrates the impact such modifications can have on complex real-world software. 
    more » « less
  3. Obeid, I. ; Selesnik, I. ; Picone, J. (Ed.)
    The Neuronix high-performance computing cluster allows us to conduct extensive machine learning experiments on big data [1]. This heterogeneous cluster uses innovative scheduling technology, Slurm [2], that manages a network of CPUs and graphics processing units (GPUs). The GPU farm consists of a variety of processors ranging from low-end consumer grade devices such as the Nvidia GTX 970 to higher-end devices such as the GeForce RTX 2080. These GPUs are essential to our research since they allow extremely compute-intensive deep learning tasks to be executed on massive data resources such as the TUH EEG Corpus [2]. We use TensorFlow [3] as the core machine learning library for our deep learning systems, and routinely employ multiple GPUs to accelerate the training process. Reproducible results are essential to machine learning research. Reproducibility in this context means the ability to replicate an existing experiment – performance metrics such as error rates should be identical and floating-point calculations should match closely. Three examples of ways we typically expect an experiment to be replicable are: (1) The same job run on the same processor should produce the same results each time it is run. (2) A job run on a CPU and GPU should produce identical results. (3) A job should produce comparable results if the data is presented in a different order. System optimization requires an ability to directly compare error rates for algorithms evaluated under comparable operating conditions. However, it is a difficult task to exactly reproduce the results for large, complex deep learning systems that often require more than a trillion calculations per experiment [5]. This is a fairly well-known issue and one we will explore in this poster. Researchers must be able to replicate results on a specific data set to establish the integrity of an implementation. They can then use that implementation as a baseline for comparison purposes. A lack of reproducibility makes it very difficult to debug algorithms and validate changes to the system. Equally important, since many results in deep learning research are dependent on the order in which the system is exposed to the data, the specific processors used, and even the order in which those processors are accessed, it becomes a challenging problem to compare two algorithms since each system must be individually optimized for a specific data set or processor. This is extremely time-consuming for algorithm research in which a single run often taxes a computing environment to its limits. Well-known techniques such as cross-validation [5,6] can be used to mitigate these effects, but this is also computationally expensive. These issues are further compounded by the fact that most deep learning algorithms are susceptible to the way computational noise propagates through the system. GPUs are particularly notorious for this because, in a clustered environment, it becomes more difficult to control which processors are used at various points in time. Another equally frustrating issue is that upgrades to the deep learning package, such as the transition from TensorFlow v1.9 to v1.13, can also result in large fluctuations in error rates when re-running the same experiment. Since TensorFlow is constantly updating functions to support GPU use, maintaining an historical archive of experimental results that can be used to calibrate algorithm research is quite a challenge. This makes it very difficult to optimize the system or select the best configurations. The overall impact of all of these issues described above is significant as error rates can fluctuate by as much as 25% due to these types of computational issues. Cross-validation is one technique used to mitigate this, but that is expensive since you need to do multiple runs over the data, which further taxes a computing infrastructure already running at max capacity. GPUs are preferred when training a large network since these systems train at least two orders of magnitude faster than CPUs [7]. Large-scale experiments are simply not feasible without using GPUs. However, there is a tradeoff to gain this performance. Since all our GPUs use the NVIDIA CUDA® Deep Neural Network library (cuDNN) [8], a GPU-accelerated library of primitives for deep neural networks, it adds an element of randomness into the experiment. When a GPU is used to train a network in TensorFlow, it automatically searches for a cuDNN implementation. NVIDIA’s cuDNN implementation provides algorithms that increase the performance and help the model train quicker, but they are non-deterministic algorithms [9,10]. Since our networks have many complex layers, there is no easy way to avoid this randomness. Instead of comparing each epoch, we compare the average performance of the experiment because it gives us a hint of how our model is performing per experiment, and if the changes we make are efficient. In this poster, we will discuss a variety of issues related to reproducibility and introduce ways we mitigate these effects. For example, TensorFlow uses a random number generator (RNG) which is not seeded by default. TensorFlow determines the initialization point and how certain functions execute using the RNG. The solution for this is seeding all the necessary components before training the model. This forces TensorFlow to use the same initialization point and sets how certain layers work (e.g., dropout layers). However, seeding all the RNGs will not guarantee a controlled experiment. Other variables can affect the outcome of the experiment such as training using GPUs, allowing multi-threading on CPUs, using certain layers, etc. To mitigate our problems with reproducibility, we first make sure that the data is processed in the same order during training. Therefore, we save the data from the last experiment and to make sure the newer experiment follows the same order. If we allow the data to be shuffled, it can affect the performance due to how the model was exposed to the data. We also specify the float data type to be 32-bit since Python defaults to 64-bit. We try to avoid using 64-bit precision because the numbers produced by a GPU can vary significantly depending on the GPU architecture [11-13]. Controlling precision somewhat reduces differences due to computational noise even though technically it increases the amount of computational noise. We are currently developing more advanced techniques for preserving the efficiency of our training process while also maintaining the ability to reproduce models. In our poster presentation we will demonstrate these issues using some novel visualization tools, present several examples of the extent to which these issues influence research results on electroencephalography (EEG) and digital pathology experiments and introduce new ways to manage such computational issues. 
    more » « less
  4. Obeid, Iyad ; Selesnick, Ivan ; Picone, Joseph (Ed.)
    The goal of this work was to design a low-cost computing facility that can support the development of an open source digital pathology corpus containing 1M images [1]. A single image from a clinical-grade digital pathology scanner can range in size from hundreds of megabytes to five gigabytes. A 1M image database requires over a petabyte (PB) of disk space. To do meaningful work in this problem space requires a significant allocation of computing resources. The improvements and expansions to our HPC (highperformance computing) cluster, known as Neuronix [2], required to support working with digital pathology fall into two broad categories: computation and storage. To handle the increased computational burden and increase job throughput, we are using Slurm [3] as our scheduler and resource manager. For storage, we have designed and implemented a multi-layer filesystem architecture to distribute a filesystem across multiple machines. These enhancements, which are entirely based on open source software, have extended the capabilities of our cluster and increased its cost-effectiveness. Slurm has numerous features that allow it to generalize to a number of different scenarios. Among the most notable is its support for GPU (graphics processing unit) scheduling. GPUs can offer a tremendous performance increase in machine learning applications [4] and Slurm’s built-in mechanisms for handling them was a key factor in making this choice. Slurm has a general resource (GRES) mechanism that can be used to configure and enable support for resources beyond the ones provided by the traditional HPC scheduler (e.g. memory, wall-clock time), and GPUs are among the GRES types that can be supported by Slurm [5]. In addition to being able to track resources, Slurm does strict enforcement of resource allocation. This becomes very important as the computational demands of the jobs increase, so that they have all the resources they need, and that they don’t take resources from other jobs. It is a common practice among GPU-enabled frameworks to query the CUDA runtime library/drivers and iterate over the list of GPUs, attempting to establish a context on all of them. Slurm is able to affect the hardware discovery process of these jobs, which enables a number of these jobs to run alongside each other, even if the GPUs are in exclusive-process mode. To store large quantities of digital pathology slides, we developed a robust, extensible distributed storage solution. We utilized a number of open source tools to create a single filesystem, which can be mounted by any machine on the network. At the lowest layer of abstraction are the hard drives, which were split into 4 60-disk chassis, using 8TB drives. To support these disks, we have two server units, each equipped with Intel Xeon CPUs and 128GB of RAM. At the filesystem level, we have implemented a multi-layer solution that: (1) connects the disks together into a single filesystem/mountpoint using the ZFS (Zettabyte File System) [6], and (2) connects filesystems on multiple machines together to form a single mountpoint using Gluster [7]. ZFS, initially developed by Sun Microsystems, provides disk-level awareness and a filesystem which takes advantage of that awareness to provide fault tolerance. At the filesystem level, ZFS protects against data corruption and the infamous RAID write-hole bug by implementing a journaling scheme (the ZFS intent log, or ZIL) and copy-on-write functionality. Each machine (1 controller + 2 disk chassis) has its own separate ZFS filesystem. Gluster, essentially a meta-filesystem, takes each of these, and provides the means to connect them together over the network and using distributed (similar to RAID 0 but without striping individual files), and mirrored (similar to RAID 1) configurations [8]. By implementing these improvements, it has been possible to expand the storage and computational power of the Neuronix cluster arbitrarily to support the most computationally-intensive endeavors by scaling horizontally. We have greatly improved the scalability of the cluster while maintaining its excellent price/performance ratio [1]. 
    more » « less
  5. Outlier detection (OD) is a key machine learning task for finding rare and deviant data samples, with many time-critical applications such as fraud detection and intrusion detection. In this work, we propose TOD, the first tensor-based system for efficient and scalable outlier detection on distributed multi-GPU machines. A key idea behind TOD is decomposing complex OD applications into a small collection of basic tensor algebra operators. This decomposition enables TOD to accelerate OD computations by leveraging recent advances in deep learning infrastructure in both hardware and software. Moreover, to deploy memory-intensive OD applications on modern GPUs with limited on-device memory, we introduce two key techniques. First, provable quantization speeds up OD computations and reduces its memory footprint by automatically performing specific floating-point operations in lower precision while provably guaranteeing no accuracy loss. Second, to exploit the aggregated compute resources and memory capacity of multiple GPUs, we introduce automatic batching , which decomposes OD computations into small batches for both sequential execution on a single GPU and parallel execution across multiple GPUs. TOD supports a diverse set of OD algorithms. Evaluation on 11 real-world and 3 synthetic OD datasets shows that TOD is on average 10.9X faster than the leading CPU-based OD system PyOD (with a maximum speedup of 38.9X), and can handle much larger datasets than existing GPU-based OD systems. In addition, TOD allows easy integration of new OD operators, enabling fast prototyping of emerging and yet-to-discovered OD algorithms. 
    more » « less