From Laptop to Lambda: Outsourcing Everyday Jobs to Thousands of Transient Functional Containers

ATC 19

Presentation

• Occasional task that needs 10,000 cores (in the cloud)

• Many others share this dream

  • Outsourcing computation

  • Cluster-computing framework

  • Burst-parallel cloud functions

• Limited speed-ups, high costs, limited applicability

• gg: framework and a toolkit that makes it practical to outsource everyday applications using thousands of parallel threads in cloud services

• Challenges of outsourcing applications to the cloud

  • Software dependencies must be managed

    • With data flow frameworks like Spark, Hadoop, and Dryad, the software dependencies remain unmanaged

    • Need a warm cluster with everything preloaded

    • Not amenable to occasional one-off tasks

    • A 10,000-core cluster on EC2 is expensive!

    • Thunk abstractions

      • Lightweight container

      • Identifies an executable, along with its argument, environment, and input data

      • Data is named by the hash of its content

      • 

        • Full functional footprint

  • Roundtrips to the cloud hurt performance

    • Current application-specific outsourcing tools perform better over fast networks: distcc, icecc, UCop

    • The laptop is in the driver's seat! Want to take laptop out of the loop. Minimize the communication in-between

    • 
    • Containers can reference each other's outputs: linked containers (gg IR)

    • Graphs, and dynamic dependency graphs etc.

  • Cloud functions are promising, but hard to use well

    • The dream: renting a supercomputer by the second

    • Warm clusters are expensive, cold clusters are slow to start

    • 10,000 workers for 10 seconds on AWS Lambda costs ~$5!

      • PyWren, Sprocket, Serverless MapReduce

    • Using cloud functions is challenging

    • gg on Lambda

      • Faster speedup

      • Getting data to the cloud is faster (HTTP pipelining, multi-threading)

    • Many applications require inter-function communication

      • Current systems use indirect techniques such as using shared storage (e.g., S3)

    • Using off-the-shelf NAT-traversal techniques, the Lambdas can talk to each other at speed up to 600 Mbps

• Example: applications - software compilation

  • Build systems are often large and complicated; very difficult to manually rewrite them in gg IR.

  • We need a system that works with existing build systems, like make, Cmake, nija, etc.

  • Technique: model substitution - a technique to extract gg IR from existing applications

    • Idea: run the original build system, but replace every stage with a 'model' program that produces a thunk, instead of the actual output

  • gg on AWS Lambda is 2-5x faster than icecc outsourcing to a 384-core cluster

    • Performance doesn't increase as you get more cores, because the laptop becomes the bottleneck

    • And much more cheaper

  • Google Chrome: 18 mins...