Hello Michael,
If you have e.g., universities and ISPs donating resources (by running
some sort of cache/mirror servers) it would be possible to do things a
little simpler. If I understand you correctly, you could essentially
combine the donated resources with an existing video server or server
infrastructure to form a larger CDN that would help reduce the load on
the video server. How the CDN communicated internally would not matter
as long as it was able to accept normal HTTP requests from end users,
and it would be possible to do something like scenario 3 on this page:
http://www.inet.no/shepherd/p2pd/http_cdn.html
A university that wished to donate network resources would install a
cache server that could accept requests for e.g., Wikimedia video
files via HTTP and then register the IP/port number of the cache with
Wikimedia, which would list it as a mirror. When the cache received a
client request for a file it did not already have cached it would
retrieve it before sending the data to the client. Rather than using
HTTP internally, the video server and various caches could use a P2P
protocol for communication to transmit data efficiently and reduce the
load on the video server.
Additional benefits of an approach like this is that only the
requested files are copied to the caches, and end-users would not need
to install any new software; it would simply look like there were many
normal HTTP mirrors. A drawback, compared to a system where all the
caches have a copy of all the files on the video server, would
potentially be extra latency each time a cache needs to retrieve data
it does not have cached.
Having adaptive streaming would be quite interesting. As long as the
client does the adaption or stream selection, there should be no
problems with supporting this in a P2P-based system. The system would
however probably be most efficient if it supported variable block
sizes, with each chunk corresponding to one block.
Regards,
Karl-Andre' Skevik
Michael Dale <mdale(a)wikimedia.org> writes:
My quick thought would be to think how the system
could work over
vanilla http. ie clients can't really be expected to install things. If
the files are split into chunks per keyframe set, could a javascript
system drive reassembly? And or could this work in conjunction with
efforts to build adaptive streaming protocol.
http://blog.gingertech.net/2010/10/09/adaptive-http-streaming-for-open-code…
Where each chunk could potentially be pulled from different servers.
The p2p architecture shifts to a relatively small amount of medium sized
http support servers rather than large amount of p2p clients. Ie
something medium sized hosts like universities and small ISPs could
donate to, and have it work over normal http
I have also cc'ed people who worked on the p2p-next project, Arno
Bakker, Riccardo, Diego and Jan Gerber. ( You guys may want to join
Wikivideo-l list ). The p2p next group is working on a new
architecture for p2p distribution that works over UDP ( so pretty
different from the above proposal ) they presented their work at OVC.
Arno is there a paper up somewhere that outlines the architecture of the
new system your working on?
peace,
--michael
On 10/14/2010 04:42 AM, Karl-Andre' Skevik wrote:
Hello everybody,
I recently noticed that Wikipedia has starting using a P2P based
system for distributing video, and since I am working on a similar
system I contacted Michael Dale who suggested that I send an email to
this list regarding this.
The most simple way to describe the system is that it is conceptually
very similar to HTTP with P2P support; the protocol is essentially
HTTP with additional commands to handle P2P operations.
Files are stored on a server that also keeps track of other hosts
where all or parts of the content can be found. Before downloading a
file, clients will contact the server to obtain general file
information, such as file and block size. Files are divided into
blocks and clients request a list of hosts where the blocks they want
to retrieve can be found. The list might contain only a fixed set of
addresses (such as that of the server and any mirrors), or it can
include the addresses of other clients that have retrieved the blocks
previously. Clients attempt to automatically determine the best hosts
from the list to retrieve the content from and later report successful
block downloads to the server.
The system also supports caches, similar to web caches, to reduce
redundant network traffic.
More information can be found here:
http://www.inet.no/shepherd/p2pd/
The software was previously extensively tested on PlanetLab, but now
needs more testing after a major restructuring of the code. It also
needs to be ported to Windows.
It is in other words not yet ready for production usage, but it would
be interesting to discuss potential challenges with e.g., browser
integration when a P2P-based protocol is used for video streaming. Not
only will the protocol most likely not be directly supported by
browsers, but to get the most benefit from a P2P system users should
ideally continue to share the content they have downloaded after
having finished watching it. This will require some kind of server
process to continue running, perhaps independently of the browser.
Especially for infrequently accessed video files this would likely
have an impact on the degree to which using a P2P system will reduce
the server load.
The following page has links to a few documents that describe various
usage scenarios and some simple ways of integrating p2pd with a
browser:
http://www.inet.no/shepherd/p2pd/docs.html
Any thoughts on these integration issues from the perspective of
Wikipedia?
Regards,
Karl-Andre' Skevik
_______________________________________________
Wikivideo-l mailing list
Wikivideo-l(a)lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikivideo-l
_______________________________________________
Wikivideo-l mailing list
Wikivideo-l(a)lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikivideo-l LocalWords: Wikimedia HTTP