Management Overlay Networks (MON)
Since dynamic status query and control can often be finished
in a short time (e.g., minutes to hours, but not years), MON
takes a novel on-demand and no-repair approach.
An overlay is built from scratch whenever some management tasks need to
be executed. The overlay is discarded as soon as the tasks are
finished. Since each overlay is only maintained for a short time,
there is no need to have complex failure repair mechanisms. As
a result, MON is extremely simple and unlikely to exhibit
unexpected, emergent behaviors.
In the MON project, we have focused on designing new algorithms
for on-demand overlay construction to achieve high coverage,
high reliability and low response time.
We have also implemented a SQL like query language and
various management commands to make MON practically useful. MON is
currently one of the infrastructure services running on PlanetLab
and it provides the ability to dynamically query the PlanetLab status.
Below we briefly describe how to build overlays on-demand. The detailed
usage of MON is described here.
An early version of MON also supports software push. However,
we have since removed the component and focused on status query
and control.
On-demand overlay construction is very different from
existing P2P overlays, which are built and maintained
by handling individual joins and departures.
For on-demand overlays, however, the goal is to setup
an overlay to connect a set of existing nodes. As a result,
membership information is very important to the construction
algorithm. Designing an overlay construction algorithm involves the
joint design
of a membership layer and the overlay construction protocol
To achieve high coverage, we have designed
several membership layers that self-organize nodes into certain
loose form of overlay structures. The on-demand protocol can
make use of such loose structure for achieve probabilistic high
coverage and low response time. In addition, we have designed
several techniques such as incremental overlay construction,
DAG (direct acyclic graph) based opportunistic aggregation to
improve the coverage, reliability and performance of on-demand
overlays.
We note the separation of overlay construction
into a membership layer and on-demand protocol provides a lot
of freedom for novel algorithm design.
On the one extreme, the
membership layer can just use simple gossip protocols for
random membership maintenance. This provides little support
for the on-demand protocol, thus the resulting algorithm may
not have good coverage and performance. On the other extreme,
the membership layer can maintain a tree structure by itself.
This effectively becomes a persistent overlay, thus may involve complex
failure repairs. The best tradeoff is probably somewhere in between,
where the membership layer maintains some loosely
structured information, which is easy to maintain,
yet still provides useful information that the on-demand
protocol can utilize for better overlay construction.
[What's New]
[People]
[Introduction]
[MON Usage]
[Download]
[Publications]
[Related Projects]
What's New
People
Introduction
Many large scale applications (e.g., content distribution,
name service, publish/subscribe) are being deployed in wide
area environment such as the PlanetLab testbed. To effectively
manage such applications, the ability
to dynamically query and control the application status is needed, for
example, to find out all the application nodes that have
reported a particular error message in their log files, or
to restart all application nodes that are using too much
free memory. Such dynamic query focuses on unplanned
management tasks and provides access to unforeseen
data attributes. Thus it is complementary to existing
continuous monitoring systems.
MON consists of one daemon process running on each
distributed node. The daemon process has a three-layer
architecture as shown in the figure above.
The lowest layer is responsible for (partial) membership
maintenance.
It enables each node to learn about new nodes and detect
node failures. It also allows nodes to measure network delay
between themselves.
The middle layer is responsible for on-demand overlay construction.
Currently we allow the construction of overlay trees and DAGs
(direct acyclic graphs). The top layer is responsible for
management command execution.
MON Usage
MON provides a
web interface to dynamically query the PlanetLab status.
The detailed description of the SQL like query language and a
commandline monclient are available at the
MON Usage page.
Download
The MON project is available for download under the Illinois Open Source License: Download MON source code
Publications
Jin Liang, Steven Ko, Indranil Gupta and Klara Nahrstedt, "MON: On-demand Overlays for Distributed System Management", accepted to the 2nd USENIX Workshop on Real, Large Distributed Systems (WORLDS'05)
Jin Liang, Steven Ko, Indranil Gupta and Klara Nahrstedt, "MON: Management Overlay Networks for Distributed Systems", poster session of SOSP 2005.
Jin Liang, Steven Ko, Indranil Gupta and Klara Nahrstedt, "MON: Design and Implementation of Management Overlay Networks for Distributed Systems", UIUC Tech Report, April, 2005
Related Projects
Funding Agencies
The MON project is supported by the NSF under Grant number NSF ANI 03-23434 and NSF CAREER grant CNS-0448246. Any opinions,
findings, and conclusions or recommendations expressed in
this material are those of the author(s) and do not necessarily
reflect the views of the National Science Foundation or
US government.
