<div dir="ltr">We're interested in topology support in order to support placement locality (to optimize task placement in ensembles). Vish and I started talking about what would be needed a few months ago, and came up with two approaches that would work:<br>
<div> - modelling the system as a full graph (ie rich enough topology information in order to represent orthogonal concerns, like power and network, for example)</div><div> - a limited approach where location was described through a feature vector that could be used for determining group diameter, which could be in turn used to compute group affinity and dispersion. </div>
<div><br></div><div>We're also beginning to think about trying to expose network topology upwards for scheduling as well. When you are interested in full topology, you can't take any shortcuts, so I think that we're stuck with a full graph for this. </div>
<div><br></div><div>It definitely makes sense to have a well maintained, flexible single definition of this data that can be used everywhere.</div><div> -nld</div></div>