Friday, April 27, 2018

Thesis Defense: Systems Support for Intermittent Computing

Today I attended Alexei Colin's thesis defense titled, Systems Support for Intermittent Computing.

For small, embedded devices, batteries are expensive / difficult, so energy can be harvested from RF, light, temp gradients, motion, et cetera.  In such a device, this direct energy source is insufficient to power the device, so a small capacitor (or other storage medium) retains this energy until the device can be powered for a short time.  The discharge provides an intermittent period of execution before the power source drops below the threshold for execution.  Programs can be annotated with latches or other progress points, such that execution after power failure can then resume at this point after the power is again available.

To model the computation, the program will be decomposed into tasks, where each task can only transfer control to other tasks, but contains arbitrary code.  Tasks will communicate through channels.  The channels provide the memory model, such that any internal updates within the task are ultimately exposed via the channels.  However, this model while reducing the overhead required to execute the tasks, requires a greater quantity of the non-volatile memory.

How do we then get the tasks and thus the latches?  Given a control flow graph (CFG), task boundaries will need to be inserted between specific basic blocks of the graph.  The compiler can be extended to model (or receive model results) of the energy requirements for each block and thereby estimate which path segments will have sufficient energy for complete execution.  Each block actually has not a single energy, but a PDF based on possible microarchitectural effects.  Then the model combines these PDFs to compute the CDF to determine the probability that a given path will successfully execute given a specific amount of energy available.  Note, each task boundary imposes overhead both in time and therefore energy, so we want the set of task boundaries to minimize overhead, while also accounting for task failures wasting energy.  This compiler pass produces better task decompositions than are achieved via manual programmer annotations, as provided by prior work.

Other system support issues.  This system should also have dynamic ability to select the stored energy necessary for task execution.  This change first requires splitting the energy storage device into multiple banks in hardware.  Also, debugging issues in the system is difficult, particularly where the device is expecting to regularly "fail", so a new debugger was prepared that can combine the program state of traditional debuggers, while still supporting the device to be intermittent.  Such devices will also need further design for intermittent networking stacks, and then be built into a larger IoT hierarchy.

In conclusion, energy-harvesting embedded computers will form the edge of the IoT hierarchy.  And the system stack will form the basis for support.