The volume of ads ranked impacts the performance of ad ranking systems.To enhance efficiency, multi-stage ranking systems are widely studied in academia and adopted across industry.However, as large-scale deep learning recommendation models gain prevalence, resource constraints-especially CPU and GPU limitationshave become a significant bottleneck.These constraints can hinder model iteration and lead to incomplete ranking, causing regressions in user experience and ad performance.To address these issues, we analyzed ad ranking metrics and found that ad rankings for individual users remain relatively stable over short periods.Based on this insight, we introduce Negative Exclusion Filtering, a framework that optimizes the balance between ranked ad volume and computing resources.By skipping re-ranking of consistently low-ranked ads for each user request, it reduces computing cost in large-scale social media environments.

Unexpected bursty traffic brought by certain sudden events, such as news in the spotlight on a social network or discounted items on sale, can cause severe load imbalance in backend services. Migrating hot data - the standard approach to achieve load balance - meets a challenge when handling such unexpected load imbalance, because migrating data will slow down the server that is already under heavy pressure. This article proposes PostMan, an alternative approach to rapidly mitigate load imbalance for services processing small requests. Motivated by the observation that processing large packets incurs far less CPU overhead than processing small ones, PostMan deploys a number of middleboxes called helpers to assemble small packets into large ones for the heavily-loaded server. This approach essentially offloads the overhead of packet processing from the heavily-loaded server to helpers. To minimize the overhead, PostMan activates helpers on demand, only when bursty traffic is detected. The heavily-loaded server determines when clients connect/disconnect to/from helpers based on the real-time load statistics. To tolerate helper failures, PostMan can migrate connections across helpers and can ensure packet ordering despite such migration. Driven by real-world workloads, our evaluation shows that, with the help of PostMan, a Memcached server can mitigate bursty traffic within hundreds of milliseconds, while migrating data takes tens of seconds and increases the latency during migration.

Testing a scalability bottleneck requires a large system to generate sufficient load, which is usually not accessible to researchers. To address this problem, this paper extrapolates the workload to a bottleneck node. The key observation that motivates our approach is that systems at a large scale are often repeating their behaviors at small scales, by running a job more times, running more nodes of the same type, or running more iterations of the same loop. Following this observation, we record a node's workloads at small scales and extrapolate such workload at a large scale. Towards this goal, we have developed PatternMiner, a semi-automatic tool to identify how workload patterns change with scale. We have tested our method on HDFS NameNode and YARN's Resource Manager. Our evaluation shows that PatternMiner is able to predict 98% of the workloads for NameNode and 83% of the workloads for the Resource Manager. Furthermore, by utilizing the extrapolated workload, we are able to emulate a cluster of up to 60,000 nodes with only 8 physical machines to evaluate NameNode and Resource Manager.

Increasing number of MPI applications are being ported to take advantage of the compute power offered by GPUs. Data movement on GPU clusters continues to be the major bottleneck that keeps scientific applications from fully harnessing the potential of GPUs. Earlier, GPU-GPU inter-node communication has to move data from GPU memory to host memory before sending it over the network. MPI libraries like MVAPICH2 have provided solutions to alleviate this bottleneck using host-based pipelining techniques. Besides that, the newly introduced GPU Direct RDMA (GDR) is a promising solution to further solve this data movement bottleneck. However, existing design in MPI libraries applies the rendezvous protocol for all message sizes, which incurs considerable overhead for small message communications due to extra synchronization message exchange. In this paper, we propose new techniques to optimize internode GPU-to-GPU communications for small message sizes. Our designs to support the eager protocol include efficient support at both sender and receiver sides. Furthermore, we propose a new data path to provide fast copies between host and GPU memories. To the best of our knowledge, this is the first study to propose efficient designs for GPU communication for small message sizes, using eager protocol. Our experimental results demonstrate up to 59% and 63% reduction in latency for GPU-to-GPU and CPU-to-GPU point-to-point communications, respectively. These designs boost the uni-directional bandwidth by 7.3x and 1.7x, respectively. We also evaluate our proposed design with two end-applications: GPULBM and HOOMD-blue. Performance numbers on Kepler GPUs shows that, compared to the best existing GDR design, our proposed designs achieve up to 23.4% latency reduction for GPULBM and 58% increase in average TPS for HOOMD-blue, respectively.

Virtualization has become a central role in HPC Cloud due to easy management and low cost of computation and communication. Recently, Single Root I/O Virtualization (SR-IOV) technology has been introduced for high-performance interconnects such as InfiniBand and can attain near to native performance for inter-node communication. However, the SR-IOV scheme lacks locality aware communication support, which leads to performance overheads for inter-VM communication within a same physical node. To address this issue, this paper first proposes a high performance design of MPI library over SR-IOV enabled InfiniBand clusters by dynamically detecting VM locality and coordinating data movements between SR-IOV and Inter-VM shared memory (IVShmem) channels. Through our proposed design, MPI applications running in virtualized mode can achieve efficient locality-aware communication on SR-IOV enabled InfiniBand clusters. In addition, we optimize communications in IVShmem and SR-IOV channels by analyzing the performance impact of core mechanisms and parameters inside MPI library to deliver better performance in virtual machines. Finally, we conduct comprehensive performance studies by using point-to-point and collective benchmarks, and HPC applications. Experimental evaluations show that our proposed MPI library design can significantly improve the performance for point-to-point and collective operations, and MPI applications with different InfiniBand transport protocols (RC and UD) by up to 158%, 76%, 43%, respectively, compared with SR-IOV. To the best of our knowledge, this is the first study to offer a high performance MPI library that supports efficient locality aware MPI communication over SR-IOV enabled InfiniBand clusters.

An increasing number of MPI applications are being ported to take advantage of the compute power offered by GPUs. Data movement continues to be the major bottleneck on GPU clusters, more so when data is non-contiguous, which is common in scientific applications. The existing techniques of optimizing MPI data type processing, to improve performance of non-contiguous data movement, handle only certain data patterns efficiently while incurring overheads for the others. In this paper, we first propose a set of optimized techniques to handle different MPI data types. Next, we propose a novel framework (HAND) that enables hybrid and adaptive selection among different techniques and tuning to achieve better performance with all data types. Our experimental results using the modified DDTBench suite demonstrate up to a 98% reduction in data type latency. We also apply this data type-aware design on an N-Body particle simulation application. Performance evaluation of this application on a 64 GPU cluster shows that our proposed approach can achieve up to 80% and 54% increase in performance by using struct and indexed data types compared to the existing best design. To the best of our knowledge, this is the first attempt to propose a hybrid and adaptive solution to integrate all existing schemes to optimize arbitrary non-contiguous data movement using MPI data types on GPU clusters.

Accelerating High-Performance Linkpack (HPL) on heterogeneous clusters with multi-core CPUs and GPUs has attracted a lot of attention from the High Performance Computing community. It is becoming common for large scale clusters to have GPUs on only a subset of nodes in order to limit system costs. The major challenge for HPL in this case is to efficiently take advantage of all the CPU and GPU resources available on a cluster. In this paper, we present a novel two-level workload partitioning approach for HPL that distributes workload based on the compute power of CPU/GPU nodes across the cluster. Our approach also handles multi-GPU configurations. Unlike earlier approaches for heterogeneous clusters with CPU and GPU nodes, our design takes advantage of asynchronous kernel launches and CUDA copies to overlap computation and CPU-GPU data movement. It uses techniques such as process grid reordering to reduce MPI communication/contention while ensuring load balance across nodes. Our experimental results using 32 GPU and 128 CPU nodes of Oakley, a research cluster at Ohio Supercomputer Center, shows that our proposed approach can achieve more than 80% of combined actual peak performance of CPU and GPU nodes. This provides 47% and 63% increase in the HPL performance that can be reported using only CPU nodes and only GPU nodes, respectively.