An official website of the United States government
Processing-in-memory (PIM), where compute is moved closer to memory or data, has been explored to accelerate emerging workloads. Different PIM-based systems have been announced, each offering a unique microarchitectural organization of their compute units, ranging from fixed functional units to programmable general-purpose compute cores near memory. However, one fundamental limitation of PIM is that each compute unit can only access its local memory; access to “remote” memory must occur through the host CPU – potentially limiting application performance scalability. In this work, we first characterize the scalability of real PIM architectures using the UPMEM PIM system. We analyze how the overhead of communicating through the host (instead of providing direct communication between the PIM compute units) can become a bottleneck for collective communications that are commonly used in many workloads. To overcome this inter-PIM bank communication, we propose PIMnet – a PIM interconnection network for PIM banks that provides direct connectivity between compute units and removes the overhead of communicating through the host. PIMnet exploits bandwidth parallelism where communication across the different PIM bank/chips can occur in parallel to maximize communication performance. PIMnet also matches the DRAM packaging hierarchy with a multi-tier network architecture. Unlike traditional interconnection networks, PIMnet is a PIM controlled network where communication is managed by the PIM logic, optimizing collective communications and minimizing the hardware overhead of PIMnet. Our evaluation of PIMnet shows that it provides up to 85× speedup on collective communications and achieves a 11.8× improvement on real applications compared to the baseline PIM.
more »
« less
Scalability Limitations of Processing-in-Memory using Real System Evaluations
Processing-in-memory (PIM), where the compute is moved closer to the memory or the data, has been widely explored to accelerate emerging workloads. Recently, different PIM-based systems have been announced by memory vendors to minimize data movement and improve performance as well as energy efficiency. One critical component of PIM is the large amount of compute parallelism provided across many PIM nodes'' or the compute units near the memory. In this work, we provide an extensive evaluation and analysis of real PIM systems based on UPMEM PIM. We show that while there are benefits of PIM, there are also scalability challenges and limitations as the number of PIM nodes increases. In particular, we show how collective communications that are commonly found in many kernels/workloads can be problematic for PIM systems. To evaluate the impact of collective communication in PIM architectures, we provide an in-depth analysis of two workloads on the UPMEM PIM system that utilize representative common collective communication patterns -- AllReduce and All-to-All communication. Specifically, we evaluate 1) embedding tables that are commonly used in recommendation systems that require AllReduce and 2) the Number Theoretic Transform (NTT) kernel which is a critical component of Fully Homomorphic Encryption (FHE) that requires All-to-All communication. We analyze the performance benefits of these workloads and show how they can be efficiently mapped to the PIM architecture through alternative data partitioning. However, since each PIM compute unit can only access its local memory, when communication is necessary between PIM nodes (or remote data is needed), communication between the compute units must be done through the host CPU, thereby severely hampering application performance. To increase the scalability (or applicability) of PIM to future workloads, we make the case for how future PIM architectures need efficient communication or interconnection networks between the PIM nodes that require both hardware and software support.
more »
« less
- Award ID(s):
- 2312276
- PAR ID:
- 10523945
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- Proceedings of the ACM on Measurement and Analysis of Computing Systems
- Volume:
- 8
- Issue:
- 1
- ISSN:
- 2476-1249
- Page Range / eLocation ID:
- 1 to 28
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
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.more » « less
-
To deliver scalable performance to large-scale scientific and data analytic applications, HPC cluster architectures adopt the distributed-memory model. The performance and scalability of parallel applications on such systems are limited by the communication cost across compute nodes. Therefore, projecting the minimum communication cost and maximum scalability of the user applications plays a critical role in assessing the benefits of porting these applications to HPC clusters as well as developing efficient distributed-memory implementations. Unfortunately, this task is extremely challenging for end users, as it requires comprehensive knowledge of the target application and hardware architecture and demands significant effort and time for manual system analysis. To streamline the process of porting user applications to HPC clusters, this paper presents CommAnalyzer, an automated framework for estimating the communication cost on distributed-memory models from sequential code. CommAnalyzer uses novel dynamic program analyses and graph algorithms to capture the inherent flow of program values (information) in sequential code to estimate the communication when this code is ported to HPC clusters. Therefore, CommAnalyzer makes it possible to project the efficiency/scalability upper-bound (i.e., Roofline) of the effective distributed-memory implementation before even developing one. The experiments with real-world, regular and irregular HPC applications demonstrate the utility of CommAnalyzer in estimating the minimum communication of sequential applications on HPC clusters. In addition, the optimized MPI+X implementations achieve more than 92% of the efficiency upper-bound across the different workloads.more » « less
-
Today’s Deep Neural Network (DNN) inference systems contain hundreds of billions of parameters, resulting in significant latency and energy overheads during inference due to frequent data transfers between compute and memory units. Processing-in-Memory (PiM) has emerged as a viable solution to tackle this problem by avoiding the expensive data movement. PiM approaches based on electrical devices suffer from throughput and energy efficiency issues. In contrast, Optically-addressed Phase Change Memory (OPCM) operates with light and achieves much higher throughput and energy efficiency compared to its electrical counterparts. This paper introduces a system-level design that takes the OPCM programming overhead into consideration, and identifies that the programming cost dominates the DNN inference on OPCM-based PiM architectures. We explore the design space of this system and identify the most energy-efficient OPCM array size and batch size. We propose a novel thresholding and reordering technique on the weight blocks to further reduce the programming overhead. Combining these optimizations, our approach achieves up to 65.2x higher throughput than existing photonic accelerators for practical DNN workloads.more » « less
-
Disaggregated memory systems achieve resource utilization efficiency and system scalability by distributing computation and memory resources into distinct pools of nodes. RDMA is an attractive solution to support high-throughput communication between different disaggregated resource pools. However, existing RDMA solutions face a dilemma: one-sided RDMA completely bypasses computation at memory nodes, but its communication takes multiple round trips; two-sided RDMA achieves one-round-trip communication but requires non-trivial computation for index lookups at memory nodes, which violates the principle of disaggregated memory. This work presents Outback, a novel indexing solution for key-value stores with a one-round-trip RDMA-based network that does not incur computation-heavy tasks at memory nodes. Outback is the first to utilize dynamic minimal perfect hashing and separates its index into two components: one memory-efficient and compute-heavy component at compute nodes and the other memory-heavy and compute-efficient component at memory nodes. We implement a prototype of Outback and evaluate its performance in a public cloud. The experimental results show that Outback achieves higher throughput than both the state-of-the-art one-sided RDMA and two-sided RDMA-based in-memory KVS by 1.06--5.03×, due to the unique strength of applying a separated perfect hashing index.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3639046
