mPL (Multilevel Placement)

The idea

The same medicine multilevel partitioning gave Fiduccia-Mattheyses, administered to analytical placement. Build a hierarchy by clustering (connectivity-driven, best-choice clustering in the mature versions); place the coarsest netlist, where any optimizer is affordable; interpolate positions down a level (declustered cells inherit and perturb their cluster's location); relax with the level's optimizer (mPL's lineage used generalized force-directed and nonlinear- programming formulations with density control); repeat to the bottom. Coarse levels make the global decisions cheaply; fine levels only polish, so total runtime grows gently while quality benefits from the long-range moves that flat descent struggles to discover.

Why the paradigm stuck

Multilevel structure buys three things at once: speed (most iterations run on small problems), stability (each level starts near a sensible solution), and escape from local minima (a cluster move at level three is a thousand-cell reorganization no flat local step would find). The mPL series carried the flag through the placement-contest era with consistently strong results, and the paradigm was absorbed wholesale: APlace and NTUplace are multilevel, and modern GPU (Graphics Processing Unit) placers keep a coarsening flavor in their multi- resolution density grids.

Where you meet it today

As the reason 'multilevel' is assumed background in every placement paper since, and directly in tools whose ancestry passes through the UCLA (University of California, Los Angeles) placement group's line.