Section: New Results

Flow systems

We are currently working with Bull SA, Manuel Selva (PhD) and Lionel Morel from the Socrates team to build a monitoring framework for dynamic data-flow system in many-core architectures. Data-flow computing models computation as a pipeline of computation units absorbing a continuous stream of data. This computing model suits application development for embedded devices such as MPEG-4 video encoders. The incoming data flow is sliced into small size token (e.g. video frames). Each time, all computational units take some tokens from their inputs and produce some tokens on their outputs. We focus [7] , [8] on a management layer for handling dynamic dataflow programs in many-core architectures, where computation units may be relocated at runtime from one core to another. The questions raised by Twitter Storm, Google Millwheel or Yahoo S4, are in essence very similar. Can our current architectures hold the information dataflow produced by users in terms of computing power and memory usage? We are currently extending these embedded results to study dataflow architectures with ATOS on flow computing inside Web browsers.

François Goichon will defend his PhD on resource access equity into best-effort operating systems such as Linux. Linux is built over a layered architecture, where each layer owns a local policy that may lead to a global policy being far from best-effort. With Guillaume Salagnac from Socrates team, we show [5] , [9] that we can develop malware user space applications exploiting embedded linux firmware and device drivers differential policy that can block other concurrent applications from accessing CPU time. When this kind of applications are installed in multi-tenant architectures as found in cloud shared space, it can slowdown the entire system. These results are interesting for Dice when considering access time in web browser. Current in-browser applications are developed in Javascript, which imposes a single threaded executed model to the developer, yet operated on a multi-core architecture. Best-effort operating systems are not the best approaches to handle flow based applications that become the norm, and we think that some small, low-level shifts, should be considered.