Deep neural network (DNN) accelerators as an example of domain-specific architecture have demonstrated great success in DNN inference. However, the architecture acceleration for equally important DNN training has not yet been fully studied. With data forward, error backward and gradient calculation, DNN training is a more complicated process with higher computation and communication intensity. Because the recent research demonstrates a diminishing specialization return, namely, “accelerator wall”, we believe that a promising approach is to explore coarse-grained parallelism among multiple performance-bounded accelerators to support DNN training. Distributing computations on multiple heterogeneous accelerators to achieve high throughput and balanced execution, however, remaining challenging. We present ACCPAR, a principled and systematic method of determining the tensor partition among heterogeneous accelerator arrays. Compared to prior empirical or unsystematic methods, ACCPAR considers the complete tensor partition space and can reveal previously unknown new parallelism configurations. ACCPAR optimizes the performance based on a cost model that takes into account both computation and communication costs of a heterogeneous execution environment. Hence, our method can avoid the drawbacks of existing approaches that use communication as a proxy of the performance. The enhanced flexibility of tensor partitioning in ACCPAR allows the flexible ratio of computations to be distributed among accelerators with different performances. The proposed search algorithm is also applicable to the emerging multi-path patterns in modern DNNs such as ResNet. We simulate ACCPAR on a heterogeneous accelerator array composed of both TPU-v2 and TPU-v3 accelerators for the training of large-scale DNN models such as Alexnet, Vgg series and Resnet series. The average performance improvements of the state-of-the-art “one weird trick” (OWT) and HYPAR, and ACCPAR, normalized to the baseline data parallelism scheme where each accelerator replicates the model and processes different input data in parallel, are 2.98×, 3.78×, and 6.30×, respectively.
more »
« less
Multi-dimensional balanced graph partitioning via projected gradient descent
Motivated by performance optimization of large-scale graph processing systems that distribute the graph across multiple machines, we consider the balanced graph partitioning problem. Compared to most of the previous work, we study the multi-dimensional variant in which balance according to multiple weight functions is required. As we demonstrate by experimental evaluation, such multi-dimensional balance is essential for achieving performance improvements for typical distributed graph processing workloads. We propose a new scalable technique for the multidimensional balanced graph partitioning problem. It is based on applying randomized projected gradient descent to a non-convex continuous relaxation of the objective. We show how to implement the new algorithm efficiently in both theory and practice utilizing various approaches for the projection step. Experiments with large-scale graphs containing up to hundreds of billions of edges indicate that our algorithm has superior performance compared to the state of the art.
more »
« less
- Award ID(s):
- 1657477
- NSF-PAR ID:
- 10303543
- Date Published:
- Journal Name:
- Proceedings of the VLDB Endowment
- Volume:
- 12
- Issue:
- 8
- ISSN:
- 2150-8097
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Calculation of many-body correlation functions is one of the critical kernels utilized in many scientific computing areas, especially in Lattice Quantum Chromodynamics (Lattice QCD). It is formalized as a sum of a large number of contraction terms each of which can be represented by a graph consisting of vertices describing quarks inside a hadron node and edges designating quark propagations at specific time intervals. Due to its computation- and memory-intensive nature, real-world physics systems (e.g., multi-meson or multi-baryon systems) explored by Lattice QCD prefer to leverage multi-GPUs. Different from general graph processing, many-body correlation function calculations show two specific features: a large number of computation-/data-intensive kernels and frequently repeated appearances of original and intermediate data. The former results in expensive memory operations such as tensor movements and evictions. The latter offers data reuse opportunities to mitigate the data-intensive nature of many-body correlation function calculations. However, existing graph-based multi-GPU schedulers cannot capture these data-centric features, thus resulting in a sub-optimal performance for many-body correlation function calculations. To address this issue, this paper presents a multi-GPU scheduling framework, MICCO, to accelerate contractions for correlation functions particularly by taking the data dimension (e.g., data reuse and data eviction) into account. This work first performs a comprehensive study on the interplay of data reuse and load balance, and designs two new concepts: local reuse pattern and reuse bound to study the opportunity of achieving the optimal trade-off between them. Based on this study, MICCO proposes a heuristic scheduling algorithm and a machine-learning-based regression model to generate the optimal setting of reuse bounds. Specifically, MICCO is integrated into a real-world Lattice QCD system, Redstar, for the first time running on multiple GPUs. The evaluation demonstrates MICCO outperforms other state-of-art works, achieving up to 2.25× speedup in synthesized datasets, and 1.49× speedup in real-world correlation functions.more » « less
-
Balanced graph partitioning is a critical step for many large-scale distributed computations with relational data. As graph datasets have grown in size and density, a range of highly-scalable balanced partitioning algorithms have appeared to meet varied demands across different domains. As the starting point for the present work, we observe that two recently introduced families of iterative partitioners---those based on restreaming and those based on balanced label propagation (including Facebook's Social Hash Partitioner)---can be viewed through a common modular framework of design decisions. With the help of this modular perspective, we find that a key combination of design decisions leads to a novel family of algorithms with notably better empirical performance than any existing highly-scalable algorithm on a broad range of real-world graphs. The resulting prioritized restreaming algorithms employ a constraint management strategy based on multiplicative weights, borrowed from the restreaming literature, while adopting notions of priority from balanced label propagation to optimize the ordering of the streaming process. Our experimental results consider a range of stream orders, where a dynamic ordering based on what we call ambivalence is broadly the most performative in terms of the cut quality of the resulting balanced partitions, with a static ordering based on degree being nearly as good.more » « less
-
null (Ed.)Triangle enumeration is a fundamental problem in large-scale graph analysis. For instance, triangles are used to solve practical problems like community detection and spam filtering. On the other hand, there is a large amount of data stored on database management systems (DBMSs), which can be modeled and analyzed as graphs. Alternatively, graph data can be quickly loaded into a DBMS. Our paper shows how to adapt and optimize a randomized distributed triangle enumeration algorithm with SQL queries, which is a significantly different approach from programming graph algorithms in traditional languages such as Python or C++. We choose a parallel columnar DBMS given its fast query processing, but our solution should work for a row DBMS as well. Our randomized solution provides a balanced workload for parallel query processing, being robust to the existence of skewed degree vertices. We experimentally prove our solution ensures a balanced data distribution, and hence workload, among machines. The key idea behind the algorithm is to evenly partition all possible triplets of vertices among machines, sending edges that may form a triangle to a proxy machine; this edge redistribution eliminates shuffling edges during join computation and therefore triangle enumeration becomes local and fully parallel. In summary, our algorithm exhibits linear speedup with large graphs, including graphs that have high skewness in vertex degree distributions.more » « less
-
Due to the developments of topographic techniques, clear satellite imagery, and various means for collecting information, geospatial datasets are growing in volume, complexity, and heterogeneity. For efficient execution of spatial computations and analytics on large spatial data sets, parallel processing is required. To exploit fine-grained parallel processing in large scale compute clusters, partitioning in a load-balanced way is necessary for skewed datasets. In this work, we focus on spatial join operation where the inputs are two layers of geospatial data. Our partitioning method for spatial join uses Adaptive Partitioning (ADP) technique, which is based on Quadtree partitioning. Unlike existing partitioning techniques, ADP partitions the spatial join workload instead of partitioning the individual datasets separately to provide better load-balancing. Based on our experimental evaluation, ADP partitions spatial data in a more balanced way than Quadtree partitioning and Uniform grid partitioning. ADP uses an output-sensitive duplication avoidance technique which minimizes duplication of geometries that are not part of spatial join output. In a distributed memory environment, this technique can reduce data communication and storage requirements compared to traditional methods. To improve the performance of ADP, an MPI+Threads based parallelization is presented. With ParADP, a pair of real world datasets, one with 717 million polylines and another with 10 million polygons, is partitioned into 65,536 grid cells within 7 seconds. ParADP performs well with both good weak scaling up to 4,032 CPU cores and good strong scaling up to 4,032 CPU cores.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.14778/3324301.3324307
