Skip to main content

Labios Project logo

LABIOS: A Distributed Label-Based I/O System

GRC-LEDFUNDED

Introduction​

HPC and Big Data environments have diverged over the years, resulting in diverging and even conflicting I/O requirements.

  • Consistency: Strong vs. Eventual
  • File Access: Shared vs. Independent
  • Namespace: Hierarchical vs. Flat
  • Hardware: Specialized vs. Commodity

Addressing these challenges is vital to HPC + Big Data Convergence, as it enables colocating conflicting I/O workloads on the same cluster without sacrificing performance through storage bridging; resource heterogeneity and data provisioning to support transparent data operations and conversions for complex hyperconverged workloads; storage malleability to improve resource utilization and job throughput in multi-tenant scenarios.

Approach​

  • Data Model

    • Labels, tuple of multiple operations & a pointer to input data
    • Storage-independent expression of application's intent

  • Label Manager

    • Build multiple labels based on the request characteristics
    • Splits or combines labels to optimal I/O sizes

  • Content Manager

    • Data labels are temporarily stored here for async data placement and computations
    • Represented as a key-value store

  • Label Dispatcher

    • Dispatches labels to workers
    • Supports various scheduling policies
    • Reorder labels while considering consistency

Use Cases​

Labios deployed for IO Accelerator

  • Labios for I/O acceleration
    • Fast distributed cache for temporary I/O
    • Ideal for Hadoop workloads with node-local I/O

Labios deployed for IO forwarding

  • Labios for I/O forwarding
    • Ideal for asynchronous I/O
    • Apps pass data to Labios
    • Labios interacts with remote storage
    • Scalability limited to the size of I/O forwarding layer

Labios deployed for IO buffering

  • Labios for I/O buffering
    • Fast temporary storage for persistent I/O
    • Data sharing between programs
    • In-situ visualization and analysis
    • Deep learning training pipelines

Labios deployed for remote storage

  • Labios as remote storage
    • Fast permanent storage
    • Transparently support storage hierarchies
    • Improved resource utilization and energy due to storage malleability
    • Opportunities for live reconfiguration, crash-restart, and code upgrades

Preliminary Results​

A bar chart (X-axis: System[Hadoop/Labios]-Device[Disk/Memory]-Configurations[baseline/Local/Remote]) displaying the count of various memory types, considering Storage Bridging on Map-Reduce workload with 3072 processes at 32MB/proc.

Storage Bridging on Map-Reduce workload, 3072 processes at 32MB/proc

Bar graph, x-axis is number of processes (384, 768, 1536, 3072), y-axis is the overall time in seconds ranging from 0 to 1800 with a step of 200. A bar chart displaying the count of various simulation types, including Storage Bridging on Map-Reduce workload with 3072 processes at 32MB/proc.

Resource Heterogeneity on HACC workload, 3072 processes, 16 timesteps

Key Takeaways!​

Labios allows for Data-centric system design with up to 6x boost in I/O performance and 65% reduction in execution time. This data-centric system approach can allow a more profound understanding of how data flows into the system, allowing for AI-driven I/O optimizations and data interoperability.

In collaboration with Sandia and Lawrence Livermore National Laboratories.