Scheduled Dataflow (SDF)

We have been investigating hybrid dataflow--von Neumann architectures for multithreaded systems. We have investigated the use of instruction, data, and I-structure cache memories with ETS dataflow.

 contact: kavi [at] cse [dot] unt [dot] edu 

 

Scheduled Dataflow

We are currently developing a new dataflow known as Scheduled Dataflow (SDF), which executes instructions synchronously. For the Scheduled Dataflow, we are developing various compile-time optimizations, including operand memory reuse; inplace updates for I-structure memories; split phase and non-split phase I-structure accesses; speculative pre-loading of thread contexts; and predicated instructions. Each PE in our SDF contains at least one unit for processing memory accesses (SPs) and at least one unit for executing operations (EPs). Our SPs and EPs are very simple, in order execution pipelines.

We have Dr. Kavi's slides from the presentation available in PDF format.

We are also exploring how our SDF can be configured into scalable, clusters, with each cluster containing a small number of SPs and EPs.

Most recently we are exploring Thread Level Speculation within the context of our SDF.

We have developed a formal model based on dataflow graphs that can be used for the specification and analysis of concurrent processing systems.

Uninterpreted dataflow graphs and stochastic dataflow graphs are isomorphic to Petri nets. We have developed the necessary formalisms to prove the isomorphism and also developed analysis techniques for directly analyzing stochastic dataflow graphs. We have developed approximation techniques for analysis using graph reductions and "nearly" completely decomposable Markov processes.

We have also developed formal theory based on first-order logic for the purpose of verifying logical specification of systems using our dataflow graphs.

Here you can find a detailed description of the Instructions for our Scheduled Dataflow Architecture in PDF format.

Here is a set of how-to instructions for SDF.

We also have a list of publications related to Scheduled Dataflow available.

Dataflow Graph Models

We have developed a formal model based on dataflow graphs that can be used for the specification and analysis of concurrent processing systems.

Uninterpreted dataflow graphs and stochastic dataflow graphs are isomorphic to Petri nets. We have developed the necessary formalisms to prove the isomorphism and also developed analysis techniques for directly analyzing stochastic dataflow graphs. We have developed approximation techniques for analysis using graph reductions and "nearly" completely decomposable Markov processes.

We have developed formal theory based on first-order logic for the purpose of verifying logical specification of systems using our dataflow graphs.

We have a list of publications related to Dataflow Graph Models available.