reef-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dhruv Mahajan <dhruv.maha...@gmail.com>
Subject Re: Understanding GroupCommunication terms
Date Thu, 16 Jun 2016 18:47:53 GMT
Hi

Please find answers inline.

>
> 1. What is the relationship between group, operator, Node, and Task?
>

A group consists of set of nodes/ evaluators that need to communicated with
each other. An operator like Broadast, Reduce, AllReduce specifies the
semantics of communication. Each node might be part of multiple groups.  A
task can call these operators to get and send the values. So in that sense
when a task uses these operators task and node are sort of same thing.


> 2. Is `Node` a logical or a physical concept? i.e. can a REEF Task
> contain multiple Nodes? My understanding is that it is a logical
> concept.
>
Node is a logical concept. In general, One REEF task has one node only. In
my mind I cannot think of case where it would be otherwise.


> 3. What does an `OperatorTopology` represent? To me it seems like it
> represents a logical Node.
>
For doing these communications nodes are organized in particular topology
(flat where master connects to all other nodes, Binary tree: where nodes
are organized in a tree fashion, Binomial topology and so on). Each node
needs to communicate only to its children and parents during any operator
call (broadcast etc.). The OperatorTopology class defines the local
topology for the node (children and parent) and then provide lower level
functions to send and receive messages, do reduce on output of children
etc. Each operator has its own OperatorTopology class even though they
belong to same group (ideally we need only one OperatorTopology per group)


> 4. In `GroupCommunicationMessage`, what is are `sourceId` and
> `destId`? Are they a logical or a physical concept? Do they map to
> Nodes, Tasks, or operators?
>
Let me get back after looking at the code :)



> 5. Can multiple operators reside on the same REEF Task? Multiple nodes?
> Multiple operators can reside on same REEF task. In ML its sort of common
> to do broadcast followed by reduce. Multiple groups also can reside on same
> task. Node and REEf-Task are one and the same thing. As I said earlier, I
> cannot think od the case when they will not be.




> Thanks!
> Andrew
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message