Dirk Kutscher

Personal web page

Distributed Computing in Information-Centric Networking

without comments

This is an introduction to our paper:

Distributed computing is the basis for all relevant applications on the Internet. Based on well-established principles, different mechanisms, implementations, and applications have been developed that form the foundation of the modern Web.

The Internet with its stateless forwarding service and end-to-endcommunication model promotes certain types of communication for distributed computing. For example, IP addresses and/or DNS names provide different means for identifying computing components. Reliable transport protocols (e.g., TCP, QUIC) promote interconnecting modules. Communication patterns such as REST and protocol implementations such as HTTP enable certain types of distributed computing interactions, and security frameworks such as TLS and the web PKI constrain the use of public-key cryptography for different security functions.

From Distributed Computing...

Distributed computing has different facets, for example, client-server computing, web services, stream processing, distributed consensus systems, and Turing-complete distributed computing platforms. There are also different perspectives on how distributed computing should be implemented on servers and network platforms, a research area that we refer to as Computing in the Network. Active Networking, one of the earliest works on computing in the network, intended to inject programmability and customization of data packets in the network itself; however, security and complexity considerations proved to be major limiting factors, preventing its wider deployment.

Dataplane programmability refers to the ability to program behavior, including application logic, on network elements and SmartNICs, thus enabling some form in-network computing. Alternatively, different types of server platforms and light-weight execution environments are enabling other forms of distributing computation in networked systems, such as architectural patterns, such as edge computing.

... To Computing in the Network

With currently available Internet technologies, we can observe a relatively succinct layering of networking and distributed computing, i.e., distributed computing is typically implemented in overlays with Content Distribution Networks (CDNs) being prominent and ubiquitous example. Recently, there has been growing interest in revisiting this relationship, for example by the IRTF Computing in the NetworkResearch Group (COINRG) – motivated by advances in network and server platforms, e.g., through the development of programmable data plane platforms and the development of different types of distributed computing frameworks, e.g., stream processing and microservice frameworks.

This is also motivated by the recent development of new distributed computing applications such as distributed machine learning (ML), and emerging new applications such as Metaverse suggest new levels of scale in terms of data volume for distributed computing and the pervasiveness of distributed computing tasks in such systems. There are two research questions that stem from these developments:

  1. How can we build distributed computing systems in the network that can leverage the on-path location of compute functions, e.g., optimally aligning stream processing topologies with networked computing platform topologies?

  2. How can the network support distributed computing in general, so that the design and operation of such systems can be simplified, but also so that different optimizations can be achieved to improve performance and robustness?

Issues in Legacy Distributed Computing

Although there are many distributed computing applications, it is also worth noting that there are many limitations and performance issues. Factors such as network latency, data skew, checkpoint overhead, back pressure, garbage collection overhead, and issues related to performance, memory management, and serialization and deserialization overhead can all influence the efficiency. Various optimization techniques can be implemented to alleviate these issues, including memory adjustment, refining the checkpointing process, and adopting efficient data structures and algorithms.

Some performance problems and complexity issues stem from the overlay nature of current systems and their way of achieving the above-mentioned mechanisms with temporary solutions based on TCP/IP and associated protocols such as DNS. For example, Network Service Mesh has been characterized as architecturally complex because of the so-called sidecar approaches and their implementation problems.

In systems that are layered on top of HTTP or TCP (or QUIC), compute nodes typically cannot assess the network performance directly – only indirectly through observed throughput and buffer under-runs. Information-centric data-flow systems, such as IceFlow, intend to provide better visibility and thus better joint optimization potential by more direct access to data-oriented communication resources. Then, some coordination tasks that are based on exchanging updates of shared application state can be elegantly mapped to named data publication in a hierarchical namespace, as the different dataset synchronization (Sync) protocols in NDN demonstrated.

Information-Centric Distributed Computing

In our paper on Distributed Computing in ICN at ACM ICN-2023, we focus on distributed computing and on how information-centricity in the network and application layer can support the development and operation of such systems. The rich set of distributed computing systems in ICN suggests that ICN provides some benefits for distributed computing that could offer advantages such as better performance, security, and productivity when building corresponding applications.

ICN with its data-oriented operation and generally more powerful forwarding layer provides an attractive platform for distributed computing. Several different distributed computing protocols and systems have been proposed for ICN, with different feature sets and different technical approaches, including Remote Method Invocation (RMI) as an interaction model as well as more comprehensive distributed computing platforms. RMI systems such as RICE leverage the fundamental named-based forwarding service in ICN systems and map requests to Interest messages and method names to content names (although the actual implementation is more intricate). Method parameters and results are also represented as content objects, which provides an elegant platform for such interactions.

ICN generally attempts to provide a more useful service to data-oriented applications but can also be leveraged to support distributed computing specifically.

Names

Accessing named data in the network as a native service can remove the need for mapping application logic identifiers such as function names to network and process identifiers (IP addresses, port numbers), thus simplifying implementation and run-time operation, as demonstrated by systems such as Named Function Networking (NFN), RICE, and IceFlow. It is worth noting that, although ICN does not generally require an explicit mapping of names to other domain identifiers, such networks require suitable forwarding state, e.g., obtained from configuration, dynamic learning, or routing.

Data-orientedness

ICN's notion of immutable data with strong name-content binding through cryptographic signatures and hashes seems to be conducive to many distributed computing scenarios, as both static data objects and dynamic computation results in those systems such as input parameters and result values can be directly sent as ICN data objects. NFN has first demonstrated this.

Securing distributed computing could be supported better in so far as ICN does not require additional dependencies on public-key or pipe securing infrastructure, as keys and certificates are simply named data objects and centralized trust anchors are not necessarily needed. Larger data collections can be aggregated and re-purposed by manifests (FLIC), enabling "small" and "big data" computing in one single framework that is congruent to the packet-level communication in a network. IceFlow uses such an aggregation approach to share identical stream processing results objects in multiple consumer contexts.

Data-orientedness eliminates the need for connections; even reliable communication in ICN is completely data-oriented. If higher-layer (distributed computing) transactions can be mapped to the network layer data retrieval, then server complexity can be reduced (no need to maintain several connections), and consumers get direct visibility into network performance. This can enable performance optimizations, such as linking network and computing flow control loops (one realization of joint optimization), as showed by IceFlow.

Location independence and data sharing

Embracing the principle of accessing named and authenticated data also enables location independence, i.e., corresponding data can be obtained from any place in the network, such as replication points (repos) and caches. This fundamentally enables better multi-source/path capabilities as well as data sharing, i.e., multiple data retrieval operations for one named data object by different consumers can potentially be completed by a cache, repo, or peer in the network.

Stateful Forwarding

ICN provides stateful, symmetric forwarding, which enables general performance optimizations such as in-network retransmissions, more control over multipath forwarding, and load balancing. This concept could be extended to support distributed computing specifically, for example, if load balancing is performed based on RTT observations for idempotent remote-method invocations.

More Networking, less Management

The combination of data-oriented, connection-less operation, and stateful (more powerful) forwarding in ICN shifts functionality from management and orchestration layers (back) to the network layer, which can enable complexity reduction, which can be especially pronounced in distributed computing. For example, legacy stream processing and service mesh platforms typically must manage connectivity between deployment units (pods in Kubernetes). In Apache Flink, a central orchestrator manages the connections between task managers (node agents). Systems such as IceFlow have demonstrated a more self-organized and decentralized stream-processing approach, and the presented principles are applicable to other forms of distributed computing.

In summary, we can observe that ICN's general approach of having the network providing a more natural (data retrieval) platform for applications benefits distributed computing in similar ways as it benefits other applications. One particularly promising approach is the elimination of layer barriers, which enables certain optimizations.

In addition to NFN, there are other approaches that jointly optimize the utilization of network and computing resources to provide network service mesh-like platforms, such as edge intelligence using federated learning, advanced CDNs where nodes can dynamically adapt to user demands according to content popularity, such as iCDN and OpenCDN, and general computing systems, such as Compute-First Networking, IceFlow, and ICedge.

Our paper on Distributed Computing in ICN at ACM ICN-2023 provides a comprehensive analysis and understanding of distributed computing systems in ICN, based on a survey of more than 50 papers. Naturally, these different efforts cannot be directly compared due to their difference in nature. We categorized different ICN distributed computing systems, and individual approaches and highlighted their specific properties.

The scope of this study is technologies for ICN-enabled distributed computing. Specifically, we divide the different approaches into four categories, as shown in the figure above: enablers, protocols, orchestration, and applications. The contributions of this study are as follows:

  1. A discussion of the benefits and challenges of distributed computing in ICN.
  2. A categorization of different proposed distributed computing systems in ICN.
  3. A discussion of lessons learned from these systems.
  4. A discussion of existing challenges and promising directions for future work.

Recent Research on Distributed Computing in ICN

I am providing some pointers to my previous research on distributed computing in ICN below.

The paper that has led to this article:

Current work in the Computing in the Network Research Group of the IRTF:

  • Dirk Kutscher, Teemu Kärkkäinen, Jörg Ott; Directions for Computing in the Network; Internet Draft draft-irtf-coinrg-dir-00, Work in Progress; August 2023

Reflexive Forwarding and Remote Method Invocation

Providing a unified remote computation capability in ICN presents some unique challenges, among which are timer management, client authorization, and binding to state held by servers, while maintaining the advantages of ICN protocol designs like CCN and NDN. In the RICE work,we developed a unified approach to remote function invocation in ICN that exploits the attractive ICN properties of name-based routing, receiver-driven flow and congestion control, flow balance, and object-oriented security while presenting a natural programming model to the application developer. The RICE protocol is leveraging an ICN extension called Reflexive Forwarding that provides ICN-idiomatic method parameter transmission.

Distributed Computing Frameworks

Leveraging RICE as a mechanism, we have developed Compute-First Networking (CFN) in ICN, a Turing-complete distributed computing platform. IceFlow is a proposal for Dataflow in ICN in a decentralized manner.

Applications

Based on Reflexive Forwarding, we have developed a concept for RESTful ICN that leverages CCNx key exchange for setting up security contexts and keys that could then be used for secure, data-oriented REST-like communication.

Delay-Tolerant LoRa leveraged Reflexive Forwarding to enable constrained LoRa nodes to "phone home" when they want to transmit data, thus enabling new ways (without central network and application servers) for connecting LoRa networks to the Internet.

Written by dkutscher

September 19th, 2023 at 3:47 am