Enrico and Dave,
Good discussions. Here is my $.01 input on data centers/controlled
environments.
Two examples came to mind:
- A datacenter may consist of multiple frameworks to schedule
applications. A
popular example is the Map-Reduce (Hadoop) framework. Preliminary network
aware capabilities exist; for example, see Rack Awareness
(http://hadoop.apache.org/common/docs/current/hdfs_user_guide.html#Rack+Awareness).
The document
(https://issues.apache.org/jira/secure/attachment/12345251/Rack_aware_HDFS_proposal.pdf)
discusses distances among nodes in simple settings as well.
- In
(http://static.usenix.org/events/hotice11/tech/full_papers/Wang_Ye.pdf),
motivated by the need of MS Lync, the framework aggregates information
beyond
network-centric but also app level information to provide an enterprise
information
portal.
For both settings, I feel that more abstract and more aggregated
presentation of
information, say using ALTO, can be valuable/preferred.
Richard
On 2/27/12 4:05 AM, Enrico Marocco wrote:
Thanks Dave,
this is indeed the kind of feedback the draft was intended to trigger.
Proposals that have been presented in the past few years for extending
the ALTO protocol on different axis, to make it suitable for carrying
different types of information and for doing that on a more dynamic
timescale, seem to indicate that there is some sort of gap in the
provisioning of network information to applications that tools currently
available have failed to fill. If that is the case, the discussion we
want to have needs to deepen in the whys and hows those tools do not
satisfy requirements of such applications -- I'm certainly not the most
expert in the area and will leave the question to the proponents of the
extensions, but I guess the answer may be related to the intrinsic
complexity of dealing with low-level information at an application level
-- and why a more abstract presentation of such information (through
ALTO or via other means) may have better chance.
You make a good point that for CDN and datacenter applications running
in controlled environment SNMP monitoring and BGP sniffing can provide
all the information such applications may possibly need. If the
incentive they provide is enough to motivate their implementation is a
key topic of discussion we want to have to identify the way forward for
ALTO.
Enrico
On 2/23/12 5:09 PM, David Harrington wrote:
Hi,
AD-hat-off ...
I am not very convinced this is a set of problems that need ALTO solutions.
When dealing with P2P scenarios, ALTO is important because endpoints for
a large amount of P2P are "unmanaged" - they are typically home users
sharing files with other home users. Home users typically do not
use/monitor protocols such as BGP, ISIS, SNMP, Conex, ECN. Frequently
consumer equipment doesn't make these protocols available/accessible to
end-users.
The information about the network, like server load, link status,
bandwidth availability, is not something the network providers
necessarily want to share. Network operators should be concerned about
sharing with anonymous users, who might well be interested in
maliciously attacking the network environment.
Data centers and CDNs typically are "managed" environments, and the
file-sharing/load-balancing/congestion control protocols are for use
within the administrative domain by the operators of the data centers or
CDNs (or between "peered" environments, where there is a certain level
of trust). These environments typically have access to protocols such as
SNMP and BGP, and how the network is "tweaked" to accommodate dynamic
traffics patterns is the business of the network provider, using
specialty applications to adapt the network at the lower layers.
Operators and their OAM protocols monitor traffic load and can set
policies to balance the load/adjust the forwarding rules as needed to
compensate for congestion, and so on. Applications running on end-hosts
do not have enough knowledge of the complete network traffic, and are in
a bad position to make policy decisions about load balancing across
servers based on bandwidth availability or server load or memory usage.
I understand that there is a need for communications between layer 7
applications and the underlying layer 4,3,and 2 functionality.There are
already protocols available that allow applications to inform the lower
layers of the network what type of traffic they plan to introduce to the
network, and the qualities of the service they prefer for their traffic.
Applications can already make use of some of the existing standards for
this purpose. Users probably do not have authorization to affect the
policy; they can request QoS within the policies configured by the
network operators. I do not see why, with few exceptions, the layer 7
application is better positioned to be the policy decision point,
especially for real-time adjustments, than the OAM functionality already
built into those lower layers, and the network provider policy
configurations. I also think that real-time adjustments by ALTO don't
seem called for, so a push model for fast dynamic updates really isn't
needed. If needed, existing push protocols such as SNMP notifications,
driven by an ALTO-SERVER-MIB, could serve this purpose just fine.
I have a concern about server-to-server sharing of information. I think
the network provider can decide which servers to share information with.
If server-to-server sharing eliminates the network provider from the
decision of whom to share data with, I consider that a problem. You, of
course, do not discuss how sharing would be done in this document, so
maybe that issue could be addressed.
Some of these ideas, such as server-to-server communications, might be
covered by a re-charter for the WG. However, developing a brand-new
protocol just for this purpose seems dubious when there are so many
existing protocols that can carry data between applications (which is
what an alto server is). I would expect that a better approach might be
to have a server and client co-resident, and using a
(server-as-client)-to-server communications.
I think you might have a hard time convincing the IESG to approve ALTO
extensions for data centers and CDNs that could already be handled by
other protocols in a managed environment.
My $.02
--
David Harrington
_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto
_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto
