Search for: All records

Low-latency and low-power edge AI is crucial for Virtual Reality and Augmented Reality applications. Recent advances demonstrate that hybrid models, combining convolution layers (CNN) and transformers (ViT), often achieve a superior accuracy/performance tradeoff on various computer vision and machine learning (ML) tasks. However, hybrid ML models can present system challenges for latency and energy efficiency due to their diverse nature in dataflow and memory access patterns. In this work, we leverage architecture heterogeneity from Neural Processing Units (NPU) and Compute-In-Memory (CIM) and explore diverse execution schemas to efficiently execute these hybrid models. We introduce H4H-NAS, a two-stage Neural Architecture Search (NAS) framework to automate the design of efficient hybrid CNN/ViT models for heterogeneous edge systems featuring both NPU and CIM. We propose a two-phase incremental supernet training in our NAS framework to resolve gradient conflicts between sampled subnets caused by different types of blocks in a hybrid model search space. Our H4H-NAS approach is also powered by a performance estimator built with NPU performance results measured on real silicon, and CIM performance based on industry IPs. H4H-NAS searches hybrid CNN-ViT models with fine granularity and achieves significant (up to 1.34%) top-1 accuracy improvement on ImageNet. Moreover, results from our algorithm/hardware co-design reveal up to 56.08% overall latency and 41.72% energy improvements by introducing heterogeneous computing over baseline solutions. Overall, our framework guides the design of hybrid network architectures and system architectures for NPU+CIM heterogeneous systems. 

Dynamic trees are a well-studied and fundamental building block of dynamic graph algorithms dating back to the seminal work of Sleator and Tarjan [STOC'81, (1981), pp. 114-122]. The problem is to maintain a tree subject to online edge insertions and deletions while answering queries about the tree, such as the heaviest weight on a path, etc. In the parallel batch-dynamic setting, the goal is to process batches of edge updates work efficiently in low (polylog n) span. Two work-efficient algorithms are known: batch-parallel Euler Tour Trees by Tseng et al. [ALENEX'19, (2019), pp. 92--106] and parallel Rake-Compress (RC) Trees by Acar et al. [ESA'20, (2020), pp. 2:1--2:23]. Both however are randomized and work efficient in expectation. Several downstream results that use these data structures (and indeed to the best of our knowledge, all known work-efficient parallel batch-dynamic graph algorithms) are therefore also randomized. In this work, we give the first deterministic work-efficient solution to the problem. Our algorithm maintains a parallel RC-Tree on n vertices subject to batches of k edge updates deterministically in worst-case O(k log(1 + n/k)) work and O(log n loglog k) span on the Common-CRCW PRAM. We also show how to improve the span of the randomized algorithm from O(log n log* n) to O(log n). Lastly, as a result of our new deterministic algorithm, we also derandomize several downstream results that make use of parallel batch-dynamic dynamic trees, previously for which the only efficient solutions were randomized. 

Recent work has shown how to augment any CAS-based concurrent data structure to support taking a snapshot of the current memory state. Taking the snapshot, as well as loads and CAS (Compare and Swap) operations, take constant time. Importantly, such snapshotting can be used to easily implement linearizable queries, such as range queries, over any part of a data structure. In this paper, we make two significant improvements over this approach. The first improvement removes a subtle and hard to reason about restriction that was needed to avoid a level of indirection on pointers. We introduce an approach, which we refer to as indirection-on-need, that removes the restriction, but yet almost always avoids indirection. The second improvement is to efficiently support snapshotting with lock-free locks. This requires supporting an idempotent CAS. We show a particularly simple solution to the problem that leverages the data structures used for snapshotting. Based on these ideas we implemented an easy-to-use C++ library, verlib, centered around a versioned pointer type. The library works with lock (standard or lock-free) and CAS based algorithms, or any combination. Converting existing concurrent data-structures to use the library takes minimal effort. We present results for experiments that use verlib to convert state-of-the-art data structures for ordered maps (a B-tree), radix-ordered maps (an ART-tree), and unordered maps (an optimized hash table) to be snapshottable. The snapshottable versions perform almost as well as the original versions and far outperform any previous implementations that support atomic range queries. 

We present a randomizedO(mlog²n) work,O(polylogn) depth parallel algorithm for minimum cut. This algorithm matches the work bounds of a recent sequential algorithm by Gawrychowski, Mozes, and Weimann [ICALP’20], and improves on the previously best parallel algorithm by Geissmann and Gianinazzi [SPAA’18], which performsO(mlog⁴n) work inO(polylogn) depth. Our algorithm makes use of three components that might be of independent interest. First, we design a parallel data structure that efficiently supports batched mixed queries and updates on trees. It generalizes and improves the work bounds of a previous data structure of Geissmann and Gianinazzi and is work efficient with respect to the best sequential algorithm. Second, we design a parallel algorithm for approximate minimum cut that improves on previous results by Karger and Motwani. We use this algorithm to give a work-efficient procedure to produce a tree packing, as in Karger’s sequential algorithm for minimum cuts. Last, we design an efficient parallel algorithm for solving the minimum 2-respecting cut problem. 

Memory latency and bandwidth are significant bottlenecks in designing in-memory indexes. Processing-in-memory (PIM), an emerging hardware design approach, alleviates this problem by embedding processors in memory modules, enabling low-latency memory access whose aggregated bandwidth scales linearly with the number of PIM modules. Despite recent work in balanced comparison-based indexes on PIM systems, building efficient tries for PIMs remains an open challenge due to tries' inherently unbalanced shape. This paper presents the PIM-trie, the first batch-parallel radix-based index for PIM systems that provides load balance and low communication under adversary-controlled workloads. We introduce trie matching-matching a query trie of a batch against the compressed data trie-as a key building block for PIM-friendly index operations. Our algorithm combines (i) hash-based comparisons for coarse-grained work distribution/elimination and (ii) bit-by-bit comparisons for fine-grained matching. Combined with other techniques (meta-block decomposition, selective recursive replication, differentiated verification), PIM-trie supports LongestCommonPrefix, Insert, and Delete in O(logP) communication rounds per batch and O(l/w) communication volume per string, where P is the number of PIM modules, l is the string length in bits, and w is the machine word size. Moreover, work and communication are load-balanced among modules whp, even under worst-case skew. 

We study the connections between sorting and the binary search tree (BST) model, with an aim towards showing that the fields are connected more deeply than is currently appreciated. While any BST can be used to sort by inserting the keys one-by-one, this is a very limited relationship and importantly says nothing about parallel sorting. We show what we believe to be the first formal relationship between the BST model and sorting. Namely, we show that a large class of sorting algorithms, which includes mergesort, quicksort, insertion sort, and almost every instance-optimal sorting algorithm, are equivalent in cost to offline BST algorithms. Our main theoretical tool is the geometric interpretation of the BST model introduced by Demaine et al. [18], which finds an equivalence between searches on a BST and point sets in the plane satisfying a certain property. To give an example of the utility of our approach, we introduce the log-interleave bound, a measure of the information-theoretic complexity of a permutation π, which is within a lg lg n multiplicative factor of a known lower bound in the BST model; we also devise a parallel sorting algorithm with polylogarithmic span that sorts a permutation π using comparisons proportional to its log-interleave bound. Our aforementioned result on sorting and offline BST algorithms can be used to show existence of an offline BST algorithm whose cost is within a constant factor of the log-interleave bound of any permutation π. 

Multiversioning is widely used in databases, transactional memory, and concurrent data structures. It can be used to support read-only transactions that appear atomic in the presence of concurrent update operations. Any system that maintains multiple versions of each object needs a way of efficiently reclaiming them.We experimentally compare various existing reclamation techniques by applying them to a multiversion tree and a multiversion hash table. Using insights from these experiments, we develop two new multiversion garbage collection (MVGC) techniques. These techniques use two novel concurrent version list data structures. Our experimental evaluation shows that our fastest technique is competitive with the fastest existing MVGC techniques, while using significantly less space on some workloads. Our new techniques provide strong theoretical bounds, especially on space usage. These bounds ensure that the schemes have consistent performance, avoiding the very high worst-case space usage of other techniques. 

Federated Learning (FL) under distributed concept drift is a largely unexplored area. Although concept drift is itself a well-studied phenomenon, it poses particular challenges for FL, because drifts arise staggered in time and space (across clients). Our work is the first to explicitly study data heterogeneity in both dimensions. We first demonstrate that prior solutions to drift adaptation, with their single global model, are ill-suited to staggered drifts, necessitating multiple-model solutions. We identify the problem of drift adaptation as a time-varying clustering problem, and we propose two new clustering algorithms for reacting to drifts based on local drift detection and hierarchical clustering. Empirical evaluation shows that our solutions achieve significantly higher accuracy than existing baselines, and are comparable to an idealized algorithm with oracle knowledge of the ground-truth clustering of clients to concepts at each time step. 

The performance of today's in-memory indexes is bottlenecked by the memory latency/bandwidth wall. Processing-in-memory (PIM) is an emerging approach that potentially mitigates this bottleneck, by enabling low-latency memory access whose aggregate memory bandwidth scales with the number of PIM nodes. There is an inherent tension, however, between minimizing inter-node communication and achieving load balance in PIM systems, in the presence of workload skew. This paper presents PIM-tree , an ordered index for PIM systems that achieves both low communication and high load balance, regardless of the degree of skew in data and queries. Our skew-resistant index is based on a novel division of labor between the host CPU and PIM nodes, which leverages the strengths of each. We introduce push-pull search , which dynamically decides whether to push queries to a PIM-tree node or pull the node's keys back to the CPU based on workload skew. Combined with other PIM-friendly optimizations ( shadow subtrees and chunked skip lists ), our PIM-tree provides high-throughput, (guaranteed) low communication, and (guaranteed) high load balance, for batches of point queries, updates, and range scans. We implement PIM-tree, in addition to prior proposed PIM indexes, on the latest PIM system from UPMEM, with 32 CPU cores and 2048 PIM nodes. On workloads with 500 million keys and batches of 1 million queries, the throughput using PIM-trees is up to 69.7X and 59.1x higher than the two best prior PIM-based methods. As far as we know these are the first implementations of an ordered index on a real PIM system.