CONtrol : Saving the Containers One at a Time

Shared-Storage Setup

Shared-Storage Architecture

Shared-storage architecture is a new paradigm in the way storage is managed in data-center. The crux of the concept is that instead of having a centralized storage system such as a NAS/SAN, we consolidate the storage present in each individual server rack. There are several advantages such as

  • Cost-effectiveness compared to NAS/SAN
  • Very power-efficient
  • Usage of off-the-shelf commodity hardware
  • Similar reliability to NAS/SAN

Challenges Associated with a Shared-Storage Setup

With new architectures come new challenges. Since the storage is distributed across multiple servers, data acess for a large chunk can require access to multiple servers. Another cruicial problem that we face is that the traffic generated by accessing storage (known as iSCSI traffic) can interfere with the existing network traffic, which we denote as storage-network conflict. Because iSCSI traffic can be very demanding, specially with reading/writing large files, it can cause high latency for the standard network traffic. The nature of interference also depends on the type of replication technique one has applied on the shared-storage architecture.

Our Solution : CONtrol

CONtrol : Setup

The way CONtrol approaches the storage-network conflict is by monitoring the QoS of every container. If the QoS of a container falls sharply, we try to adjust the network usage of the offending container. However, if this approach fails, we resort to migrating the affected container. Using this approch, we observed that we are performing quite well compared to other baselines.

Basic Methodology

PID Controller

The mathematical backend of our approach relies on something called a Proportional-integral-derivative controller or PID controller. It uses control theory techniques to maintain a constant QoS for all containers. The PID controller takes QoS metrics for all containers continuously as input and it gives the corresponding bandwidth resource requirement for the container. This bandwidth information is used to network policies to maintain the QoS of the containers.

Publications

  • CONtrol: Controlling Quality of Service of Container Networks in a Hyperconverged Platform, Sumitro Bhaumik, Kaustav Chanda, Sandip Chakraborty, 2020 IFIP Networking
  • NetStor: Network and Storage Traffic Management for Ensuring Application QoS in a Hyperconverged Data-center, Sumitro Bhaumik, Ravi Bansal, Raja Karmakar, Satish Kumar Mopur, Saikat Mukherjee, Mandar Jagannath Chitale, Sandip Chakraborty, 2020 IEEE Transactions on Cloud Computing
  • Shared Storage Software Defined Data Centers: Analyzing VM Migration Based on Application Workloads, Sumitro Bhaumik, Rohit Dhangar, Gouranga Murari, Swapnil Kumar Bishnu, Sandip Chakraborty, 2018 IEEE Global Communications Conference (GLOBECOM)