Resource disaggregation (RD) is an emerging paradigm for data center computing whereby resource-optimized servers are employed to minimize resource fragmentation and improve resource utilization. Apache Spark deployed under the RD paradigm employs a cluster of compute-optimized servers to run executors and a cluster of storage-optimized servers to host the data on HDFS. However, the network transfer from storage to compute cluster becomes a severe bottleneck for big data processing. Near-data processing (NDP) is a concept that aims to alleviate network load in such cases by offloading (or “pushing down”) some of the compute tasks to the storage cluster. Employing NDP for Spark under the RD paradigm is challenging because storage-optimized servers have limited computational resources and cannot host the entire Spark processing stack. Further, even if such a lightweight stack could be developed and deployed on the storage cluster, it is not entirely obvious which Spark queries would benefit from pushdown, and which tasks of a given query should be pushed down to storage.
This paper presents the design and implementation of a near-data processing system for Spark, SparkNDP, that aims to address the aforementioned challenges. SparkNDP works by
implementing novel NDP Spark capabilities on the storage cluster using a lightweight library of SQL operators and then developing an analytical model to help determine which Spark tasks should be pushed down to storage based on the current network and system state. Simulation and prototype implementation results show that SparkNDP can help reduce Spark query execution times when compared to both the default approach of not pushing down any tasks to storage and the outright NDP approach of pushing all tasks to storage.
more »
« less
AQUAMAN: An Analytic-Query Offloading Machine
Analytic workloads on terabyte data-sets are often run in the cloud, where application and storage servers are separate and connected via network. In order to saturate the storage bandwidth and to hide the long storage latency, such a solution requires an expensive server cluster with sufficient aggregate DRAM capacity and hardware threads. An alternative solution is to push the query computation into storage servers.
In this paper we present an in-storage Analytics QUery Offloading MAchiNe (AQUOMAN) to “offload” most SQL operators, including multi-way joins, to SSDs. AQUOMAN executes
Table Tasks, which apply a static dataflow graph of SQL operators to relational tables to produce an output table. Table Tasks use a streaming computation model, which allows AQUOMAN to process queries with a reasonable amount of DRAM for intermediate
results. AQUOMAN is a general analytic query processor, which can be integrated in the database software stack transparently. We have built a prototype of AQUOMAN in FPGAs, and using TPC-H benchmarks on 1TB data sets, shown that a single instance of 1TB AQUOMAN disk, on average, can free up 70% CPU cycles and reduce DRAM usage by 60%. One way to visualize this saving is to think that if we run queries sequentially and ignore inter-query page cache reuse, MonetDB running on a 4-core, 16GB-DRAM machine with AQUOMAN augmented SSDs performs, on average, as well as a MonetDB running on a 32-core, 128GB-DRAM machine with standard SSDs.
more »
« less
- Award ID(s):
- 1725303
- NSF-PAR ID:
- 10206787
- Date Published:
- Journal Name:
- Proceedings of the 53rd Annual IEEE/ACM International Symposium on Microarchitecture
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
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
-
null (Ed.)The Twitter-Based Knowledge Graph for Researchers project is an effort to construct a knowledge graph of computation-based tasks and corresponding outputs. It will be utilized by subject matter experts, statisticians, and developers. A knowledge graph is a directed graph of knowledge accumulated from a variety of sources. For our application, Subject Matter Experts (SMEs) are experts in their respective non-computer science fields, but are not necessarily experienced with running heavy computation on datasets. As a result, they find it difficult to generate workflows for their projects involving Twitter data and advanced analysis. Workflow management systems and libraries that facilitate computation are only practical when the users of these systems understand what analysis they need to perform. Our goal is to bridge this gap in understanding. Our queryable knowledge graph will generate a visual workflow for these experts and researchers to achieve their project goals. After meeting with our client, we established two primary deliverables. First, we needed to create an ontology of all Twitter-related information that an SME might want to answer. Secondly, we needed to build a knowledge graph based on this ontology and produce a set of APIs to trigger a set of network algorithms based on the information queried to the graph. An ontology is simply the class structure/schema for the graph. Throughout future meetings, we established some more specific additional requirements. Most importantly, the client stressed that users should be able to bring their own data and add it to our knowledge graph. As more research is completed and new technologies are released, it will be important to be able to edit and add to the knowledge graph. Next, we must be able to provide metrics about the data itself. These metrics will be useful for both our own work, and future research surrounding graph search problems and search optimization. Additionally, our system should provide users with information regarding the original domain that the algorithms and workflows were run against. That way they can choose the best workflow for their data. The project team first conducted a literature review, reading reports from the CS5604 Information Retrieval courses in 2016 and 2017 to extract information related to Twitter data and algorithms. This information was used to construct our raw ontology in Google Sheets, which contained a set of dataset-algorithm-dataset tuples. The raw ontology was then converted into nodes and edges csv files for building the knowledge graph. After implementing our original solution on a CentOS virtual machine hosted by the Virginia Tech Department of Computer Science, we transitioned our solution to Grakn, an open-source knowledge graph database that supports hypergraph functionality. When finalizing our workflow paths, we noted some nodes depended on completion of two or more inputs, representing an ”AND” edge. This phenomenon is modeled as a hyperedge with Grakn, initiating our transition from Neo4J to Grakn. Currently, our system supports queries through the console, where a user can type a Graql statement to retrieve information about data in the graph, from relationships to entities to derived rules. The user can also interact with the data via Grakn's data visualizer: Workbase. The user can enter Graql queries to visualize connections within the knowledge graph.more » « less
-
null (Ed.)Relational database management systems (RDBMS) have limited iterative processing support. Recursive queries were added to ANSI SQL, however, their semantics do not allow aggregation functions, which disqualifies their use for several applications, such as PageRank and shortest path computations. Recently, another SQL extension, iterative Common Table Expressions (CTEs), is proposed to enable users to perform general iterative computations on RDBMSs.In this work 1 , we demonstrate how iterative CTEs can be efficiently incorporated into a production RDBMS without major intrusion to the system. We have prototyped our approach on Futurewei's MPPDB, a shared nothing relational parallel database engine. The implementation is based on a functional rewrite that translates iterative CTEs to other existing SQL operators. Thus, query plans of iterative CTEs can be optimized and executed by the engine with minimal modification to the code base. We have also applied several optimizations specifically for iterative CTEs to i) minimize data movement, ii) reuse results that remain constant and iii) push down predicates to avoid unnecessary data processing. We verified our implementation through extensive experimental evaluation using real world datasets and queries. The results show the feasibility of the rewrite approach and the effectiveness of the optimizations, which improve performance by an order of magnitude in some cases.more » « less
-
Noria, first presented at OSDI 2018, is a new web application backend that delivers the same fast reads as an in-memory cache in front of the database, but without the application having to manage the cache. Even better, Noria still accepts SQL queries and allows changes to the queries without extra effort, just like a database. Noria performs well: it serves up to 14M requests per second on a single server, and supports a 5x higher load than carefully hand-tuned queries issued to MySQL. Writing web applications that tolerate high load is difficult. The reason is that the backend storage system that the application relies on—typically a relational database, like MySQL— can easily become a serious bottleneck with many clients. Each page view typically involves 10 or more database queries, which each take up CPU time on the database servers to evaluate. To avoid such slow database interactions and to reduce load on the database, applications often introduce caches (like memcached or Redis) that store already-computed query results for fast common case access. These caches, however, impose significant application complexity, because the application must query, invalidate, and maintain them. Surely there has to be a better way.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
