The MIDdleware Assurance Substrate: Enabling Strong Real-Time Guarantees in Open Systems with OpenFlow

Andrew L. King, Sanjian Chen, Insup LeeISORC 2014 · DOI

TL;DR

This paper introduces MIDAS, a middleware that leverages OpenFlow to provide strong real-time guarantees in dynamic, open systems. By controlling network configurations directly, MIDAS ensures that critical data streams meet their timing requirements, even under heavy load or when components are added or removed at runtime.

What Problem Does This Paper Address?

Developing distributed real-time and embedded (DRE) systems often relies on middleware to simplify communication and hide low-level networking details. However, achieving deterministic, real-time performance typically requires static network configurations. This approach breaks down in "open systems" where components can be added or removed dynamically, as the network resource demands are unpredictable. Traditional middleware cannot dynamically manage the network to guarantee performance in such environments. This paper addresses the challenge of providing strong, end-to-end real-time guarantees in these dynamic, open DRE systems by integrating middleware control with modern network programmability.

Methods

The core of the proposed solution is the MIDdleware Assurance Substrate (MIDAS), a publish-subscribe middleware that directly controls the network's configuration via the OpenFlow protocol.

Here's a breakdown of the key methods:

  • Quality of Service (QoS) Parameters: MIDAS defines three key QoS parameters for subscribers:
  • Maximum End-to-End Latency (Lmax): The absolute maximum time a message can take from publisher to subscriber.
  • Minimum Separation (minSep): The minimum time between consecutive messages from a publisher.
  • Maximum Separation (maxSep): The maximum time between consecutive messages from a publisher.
  • OpenFlow Integration: MIDAS uses OpenFlow to programmatically configure network switches. This allows the middleware to:
  • Control Packet Forwarding: Dictate how packets are routed.
  • Prioritize Traffic: Assign different priority levels to different data flows.
  • Rate-Limit Traffic: Prevent individual flows from overwhelming the network.
  • Global Resource Manager (GRM): This central component of MIDAS is responsible for:
  • Admission Control: Deciding whether to admit a new subscriber based on its QoS requirements and the current network state.
  • Network Configuration Generation: Creating new OpenFlow rules to accommodate new subscriptions.
  • Schedulability Analysis: Determining if the proposed network configuration will meet all QoS guarantees.
  • Network Reconfiguration: Applying the new configuration to the OpenFlow switches.
  • Flow Scheduling and Prioritization: For single-switch scenarios, MIDAS uses a rate-limiter based on minSep and message size to prevent "babbling idiot" scenarios. For multi-switch networks, it transforms the distributed scheduling problem into a sequence of local, single-switch scheduling problems by dividing the allowed latency among switches along the path. It also employs a modified Audsley's Optimal Priority Assignment (OPA) algorithm to assign priorities to flows, considering the limited number of hardware queues available on switches.

Datasets

Not discussed in the paper. The experiments were conducted on a custom test bench using off-the-shelf hardware.

Key Insights and Ideas

  • Bridging the Middleware-Network Divide: The most significant insight is that by giving middleware direct control over the network via OpenFlow, we can achieve deterministic real-time guarantees in dynamic environments where static configurations fail. This breaks down traditional silos.
  • OpenFlow as a Real-Time Enabler: OpenFlow's ability to programmatically control switch forwarding, prioritization, and rate-limiting at line rate is crucial. This allows for dynamic QoS adjustments without significant performance degradation.
  • Proactive "Babbling Idiot" Prevention: The use of minSep to configure rate-limiters at the switch ingress is an effective way to prevent a single misbehaving publisher from disrupting the entire network, a critical feature for robust real-time systems.
  • Transforming Distributed Scheduling: The approach of breaking down multi-switch scheduling into a series of single-switch problems simplifies a notoriously difficult NP-hard problem, making it computationally feasible for dynamic environments.
  • Dynamic QoS for Cyber-Physical Systems: The paper highlights the growing need for such dynamic QoS capabilities in emerging Cyber-Physical Systems (CPS), such as plug-and-play medical systems, where real-time performance is paramount and system configurations change frequently.

Key Results

The experimental evaluation on a test bench with a Pica8 OpenFlow switch demonstrated the effectiveness of MIDAS:

  • Improved Determinism: In a scenario with a high network load (815 Mbps), MIDAS-managed flows consistently met their millisecond-level timing requirements. In contrast, the same flows on a standard "best-effort" switch experienced violated latency constraints and lost nearly half of their messages.
  • Robust Fault Containment: When two publishers were deliberately configured to "babble" (saturate the network), MIDAS successfully prevented them from starving other critical flows. All messages for the remaining topic met their latency bounds, whereas on the best-effort network, all messages were dropped.
  • Message Loss Reduction: MIDAS eliminated message loss in the tested scenarios, whereas the best-effort configuration suffered significant message loss due to queue overruns.

Limitations and Future Work

  • Limitations:
    • The current prototype's QoS parameters are limited to latency and separation.
  • The schedulability analysis for multi-switch networks is based on an approximate technique.
  • The paper acknowledges that optimal distributed scheduling is NP-hard, and their approach is a practical strategy rather than an optimal one.
  • The current implementation relies on specific types of data structures for size-bound inference.
  • Future Work:
    • Exploring better scheduling and reconfiguration algorithms for OpenFlow.
  • Extending QoS support beyond timing to include other aspects like jitter or bandwidth guarantees.
  • Developing mechanisms to detect and adapt to network faults dynamically.
  • Investigating how to integrate non-publish-subscribe best-effort traffic more seamlessly.

Summary updated

BibTeX
@inproceedings{king2014middleware,
  author = {Andrew L. King and Sanjian Chen and Insup Lee},
  title = {{The MIDdleware Assurance Substrate: Enabling Strong Real-Time Guarantees in Open Systems with OpenFlow}},
  booktitle = {ISORC},
  year = {2014},
  doi = {10.1109/isorc.2014.49},
}
Figure