Concurrent Analytical Query Processing with GPUs

Wang K , Zhang K , Yuan Y , et al. 
Concurrent analytical query processing with GPUs[J]. 
Proceedings of the VLDB Endowment, 2014, 7(11):1011-1022.

ABSTRACT

• GPUs used as dedicated accelerators to process individual query
• Sharing GPUs among concurrent queries not supported
• Based on the profiling of
  • an opensource GPU query engine
  • running commonly used singlequery data warehousing workloads,
  • utilization of main GPU resources is 25%.
• low system throughput

an opensource GPU query engine一个开源的GPU查询引擎。在这上面跑普遍使用的单查询数据仓库负载

 

• proposes concurrent query execution as an solution.
• To efficiently share GPUs among concurrent queries for high throughput,
  • provide software support
  • to control and resolve resource contention
• Our solution relies on GPU query scheduling
  • device memory swapping policies
  • to address this challenge

 

• implemented a prototype system and evaluated it intensively
• results confirm the effectiveness and performance advantage of our approach
• By executing multiple GPU queries concurrently, system throughput can be improved by up to 55% compared with dedicated processing

1. INTRODUCTION

• Multitasking
  • a proven practice in computer systems to achieve high resource utilization and system throughput
• wide adoption of GPUs for analytical query processing,
• mainly used as dedicated co-processors,
  • unable to
  • support efficient executions of multiple queries concurrently.

 

• data-driven characteristics of
  • GPU operations,
  • a single query can hardly consume all GPU resources
• Dedicated query processing leads to resource underutilization,
  • limits the overall performance of the database system
• In market-critical applications
  • high-performance data warehousing
  • multi-client dataflow analysis
• large number of users demand query results simultaneously
• volume of data to be processed keeps increasing,
  • essential for user queries to make continuous progress
  • so that
  • new results can be generated constantly to
  • satisfy the goal of interactive analysis.
  • The lack of concurrent querying capability restricts
    the adoption of GPU databases in these application fields.

 

• dedicated usage of GPUs still needed for latency-
  critical queries to ensure performance isolation
• databases must support
  • concurrent multi-query execution
  • to maximize the throughput of non-latency-sensitive queries

对延迟敏感的查询，啥叫去保证性能隔离啊？？

• This consolidated usage of GPU resources enhances system efficiency and functionalities, but it makes the design of query execution engine more challenging.

啥叫This consolidated usage of GPU resources

• For maximal performance,each user query tends to reserve a large amount of GPU resources.
• Unlike CPUs
  • OS supports fine-grained context switches
  • virtual memory abstractions for resource sharing,
• current GPU hardware and system software provide
  • none of these interfaces
  • for database resource management

GPU任务不能被强占

• GPU tasks cannot be preempted once started;
  • on-demand data loading is not supported during task execution;

数据只能预先搞嘛？？

• automatic data swapping service is also missing when the device memory undergoes pressure

• without efficient coordination by the database,
• multiple GPU queries attempting to execute simultaneously cause
  • low resource usage,
  • system thrashing,
  • even query abortions,
  • degrade, overall system performance