Richard took words out of my mouth,:)
Yes, in some cases, from source to destination, there is only one hop away, or 
we only care about source endpoint address and destination endpoint address, we 
don't care about which intermediary it traverses in the path from src to 
destination, I think base protocol strong support such case.

However when we put more constraints on the path, e.g.,we need to compute an 
end to end path with these constraints. These constraints can be latency, 
packet loss, jitter.
However latency, packet loss, jitter are usually gathered from routing 
protocol, and put as per link metrics, so we need to do some aggregation when 
we apply these per link metrics
to the end to end path, e.g., per link latency, if we choose a path that 
traverses link A, link B, link C, then we can get end to end latency by 
choosing the sum of per link latency of link A, link B and link C.

Regards!
-Qin

From: [email protected] [mailto:[email protected]] On Behalf Of Y. 
Richard Yang
Sent: Monday, October 14, 2013 10:42 AM
To: IETF ALTO
Cc: Wendy Roome; [email protected]; Qin Wu
Subject: Re: [alto] ALTO Extension: A document defining multi-metrics filtering?

Let me add on: although a filtering service can be a very useful service, it 
can also be quite involved, and hence the WG may need to think through the 
issues when designing this service. For example, there are two types of use 
cases:

- end-to-end: given src set {s1, s2, s3, ..., sn} and dst set {d1, d2, d3, ..., 
dm}, return all pairs (si, dj), where si in {s1, ..., sn} and dj in {d1, ..., 
dm} such that (si, dj) satisfies the constraints;

- relay: given src s, dst d, and a relay candidate set {r1, r2, ..., rk}, 
return all of the ri such that s -> ri -> d satisfies the constraints.

Note that with relay, we will then need to worry about the "composition" 
semantics of metrics. For example, delay might be additive, loss rate (unless 
small and independent) may not be.

The relay could be even fancier (e.g., one-hop server detour such as Akamai 
one-hop detour and hence may involved two relay servers), but it may or may not 
be a good idea to go too complex, depending on if the WG can define a clean API 
(e.g., SQL select/where grammar comes to mind quickly).

Thanks!

Richard
On Sun, Oct 13, 2013 at 7:57 PM, Y. Richard Yang 
<[email protected]<mailto:[email protected]>> wrote:
Dear all,

The base ALTO protocol (http://www.ietf.org/id/draft-ietf-alto-protocol-20.txt) 
is mostly a single-cost-metric centric:

- The Cost Map filtering service uses only one cost-type (Sec. 11.3.2.3):

     object {
     CostType   cost-type;
     [JSONString constraints<0..*>;]
     [PIDFilter  pids;]
   } ReqFilteredCostMap;

   object {
     PIDName srcs<0..*>;
     PIDName dsts<0..*>;
   } PIDFilter;

...
 constraints  Defines a list of additional constraints on which
      elements of the Cost Map are returned.  This parameter MUST NOT be
      specified if this resource's capabilities (Section 11.3.2.4)
      indicate that constraint support is not available.  A constraint
      contains two entities separated by whitespace: (1) an operator,
      'gt' for greater than, 'lt' for less than, 'ge' for greater than
      or equal to, 'le' for less than or equal to, or 'eq' for equal to;
      (2) a target cost value.

- The Endpoint Cost service allows filtering (Sec. 11.5.1.3) as well, and is 
similar to Cost Map Filtering:

   object {
     CostType          cost-type;
     [JSONString       constraints<0..*>;]
     EndpointFilter    endpoints;
   } ReqEndpointCostMap;

   object {
     [TypedEndpointAddr srcs<0..*>;]
     [TypedEndpointAddr dsts<0..*>;]
   } EndpointFilter;

   constraints  Defined equivalently to the "constraints" input
      parameter of a Filtered Cost Map (see Section 11.3.2).

In other words, in the base protocol, the filtering condition and the output 
are based on the same Cost Metric. It is natural that the filtering and the 
output are based on different Cost metrics. For example, a Client asks for 
routingcost for only pairs whose latency is below a threshold (see use cases in 
http://tools.ietf.org/html/draft-randriamasy-alto-multi-cost-07).

One may argue that the filter-metric-no-equal-to-output-metric function can be 
implemented on top of the filter-and-output-using-one-metric function:

In particular, suppose the filtering is based on metrics M1 and M2, and the 
output is M3, for a set src to a set dsts. The Client can use the following 
three queries:

- Q1: Use single metric <M1, filter on M1, srcs, dsts> and obtains <srcs1, 
dsts1> in return;
- Q2: Use single metric <M2, filter on M2, srcs1, dsts1> and obtains <srcs2, 
dsts2> in return;
- Q3: Use single metric <M3, no filter, srcs2, dsts2> to get the final result.

Although this is not too bad, it is inconvenient. Note that preceding is first 
discussed by Sabine, Wendy, Nico in:
http://tools.ietf.org/html/draft-randriamasy-alto-multi-cost-07

I saw that this is also the issue discussed in
- http://tools.ietf.org/html/draft-wu-alto-json-te-01
- http://tools.ietf.org/html/draft-lee-alto-app-net-info-exchange-02

Hence, I propose that the WG extends the base protocol with this capability, as 
I see that it is quite useful. One issue is that I see three designs, and I am 
wondering if the authors are preparing on discussing their designs at the 
coming IETF, and if there is a possibility for a single, unified document, 
focusing on this issue.

Thanks a lot!

Richard

_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto

Reply via email to