Dryad: Distributed data-parallel programs from sequential building blocks. Conference Paper (PDF Available) in ACM SIGOPS Operating Systems Review. DRYAD: DISTRIBUTED DATA-. PARALLEL PROGRAMS FROM. SEQUENTIAL. BUILDING BLOCKS. Authors: Michael Isard, Mihai Budiu, Yuan Yu,. Andrew. An improvement: Ciel. Comparison. Conclusion. Dryad: Distributed Data-Parallel Programs from. Sequential Building Blocks. Course: CS
|Published (Last):||16 August 2012|
|PDF File Size:||18.46 Mb|
|ePub File Size:||13.82 Mb|
|Price:||Free* [*Free Regsitration Required]|
Copyrights for components of this work owned by others than ACM must be honored. Distributed Data-Parallel Programs from Sequential Building Blocks” Dryad is a “general-purpose, high performance distributed execution engine. It supports event-based programming style on vertex for you to write concurrent program.
It focuses more on simplicity of the programming model and reliability, efficiency and scalability of the applications while side-stepped problems like high-latency and unreliable wide-area networks, control of resources by separate federated or competing entities and ACL, etc.
Summary of “Dryad: Distributed Data-Parallel Programs from Sequential Building Blocks”
A Dryad job consists of DAG where each vertex is a program and each edge is a data channel, data channel can be shared memory, TCP pipes, or temp files. In contrast to MapReduce, Dryad doesn’t do serialization, for the vertex program’s perspective, what they see is a heap object passed from the previous vertex, which will certainly save a lot of data parsing headaches. If every vertex finishes successfully, the whole job is finished.
One caveat is you can only run 1 job in a cluster at a time, because the job manager assumes exclusive control over all computers within the cluster.
Dryad is a “general-purpose, high performance distributed execution engine. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page.
It provides task scheduling, concurrency optimization in a computer level, fault tolerance and data distribution.
To discover available resources, each computer in the cluster has a proxy daemon running, and they are registered into a central name server, they job manager queries the name server to get available computers. The application can discover the size and placement of data at run time, and buildinng the graph as the computation progresses to make efficient use of the available resources. In Dryad, a scheduler inside job manager tracks states of each vertex.
Proceedings of the Eurosys Conference March Research Areas Computer vision Systems and boocks. Abstracting with credit is permitted. One interesting property provided by Dryad is it can turn a graph G into a vertex V Gessentially similar to the composite design pattern, it improves the re-usability a lot. The runtime receives a closure from the job manager describing dryaf vertex to be run and URIs for dryac and output of the vertex.
Dryad is a general-purpose distributed execution engine for coarse-grain data-parallel applications. Dryad is designed to scale from powerful multi-core single computers, through small clusters of computers, to data centers with thousands of computers.
Dryad: distributed data-parallel programs from sequential building blocks – Dimensions
Which can potentially gives you more efficiency in a vertex execution. Concurrency arises from Dryad scheduling vertices to run simultaneously on multiple computers, or on multiple CPU cores within a computer.
Dryad achieves fault tolerance through proxy communicating with job manager, but if proxy failed, a timeout will be triggered in job manager indicating a vertex has failed. Dryad also provides a backup task mechanism when noticing a vertex has been slower than their peers, similar to the one used to MapReduce.
A Dryad job is coordinated by a process called job manager, can be either within the compute cluster or remote workstation that has access to the compute cluster. The performance is absolutely superior to a commercial database system for hand-coded read-only query. It supports vertex creation, edge creation and graph merging operations. Dryad’s DAG based data parallelization makes it more expressive for solving different large scale problems. One of the unique buolding provided by Dryad is the flexibility of fine control of an application’s data flow graph.
If any vertex failed, the job is re-run, but only to a threshold number of times, after that if the job is still failing, the entire job will blocjs failed. Dryad runs the application by executing the vertices of this graph on a set of available computers, communicating as appropriate through files, TCP pipes, and shared-memory FIFOs. Dryad also provides visualizer and web interface for monitoring of cluster states.
Dryad: Distributed Data-parallel Programs from Sequential Building Blocks – Microsoft Research
This gives programmer the opportunity to optimize trade offs between parallelism and data distribution overhead thus gives distributee performance” according to the paper. The dynamic refinement it provides also makes it efficient in a lot of cases.
The vertices provided by the application developer are quite simple and are usually written as sequential programs with no thread creation or locking.