Abstract

Presentation

Prof. Murali Annaveram (Paper Received by MICRO’17)

Introduction

Big Data Era and Memory Scaling

Storage Scaling

Fetching data from disk is a huge problem

(But many applications are) Stuck with block-based storage

Semantic Storage

  • Expose structure of data to storage controller
  • Opportunities: Underutilized processing power
  • Limitations:
    • Wimpy embedded core
    • Dynamic processor usage
    • In-SSD computation should be utilized dynamically

What kind of work can be done?

One Scenerio: Flash-Summarizer

  • Two options explored:
    • Intra-query offloading
    • Inter-query offloading

Flash Summarizer Architecture

  • Firmware architecture

Algorithmic Flow

  • Initiation: Transfer a in-SSD procedure to SSD memory
  • Computation: Return the special code if the requested page is processed in SSD
  • Finalization: Gather final in-SSD computation results and transfer to the host

Detailed View

  • Task Queue (TQ)
  • Task Controller
  • User function stacks

Application

  • Data integration: String Similarity Join: Prefix filtering and Verify

Evaluation

  • Platform: Multicore ARM processors + FPGA platform