Definition of a Delay Measurement Infrastructure and Delay-Sensitive Least-Used Policy for Reliable Server Pooling

Reliable Server Pooling defines protocols for providing highly available services. PEs of a pool may be distributed over a large geographical area, in order to provide redundancy in case of localized disasters. But the current pool policies defined in do not incorporate the fact of distances (i.e. delay) between PU and PE. This leads to a low performance for delay-sensitive applications.

This draft defines a delay measurement infrastructure for ENRP servers to add delay information into the handlespace. Furthermore, a delay-sensitive Least-Used policy is defined. Performance evaluations can be found in .

The terms are commonly identified in related work and can be found in the Aggregate Server Access Protocol and Endpoint Handlespace Redundancy Protocol Common Parameters document .

This section describes the necessary delay measurement infrastructure for the policy later defined in . It has to be provided as part of the ENRP servers.

Measuring delay for SCTP associations is easy: the SCTP protocol already calculates a smoothed round-trip time (RTT) for the primary path. This RTT only has to be queried via the standard SCTP API as defined in . By default, the calculated RTT has a small restriction: a SCTP endpoint waits up to 200ms before acknowledging a packet, in order to piggyback the acknowledgement chunk with payload data. In this case, the RTT would include this latency. Using the option SCTP_DELAYED_ACK_TIME (see ), the maximum delay before acknowledging a packet can be set to 0ms (i.e. "acknowledge as soon as possible"). After that, the RTT approximately consists of the network latency only. Then, using the RTT, the end-to-end delay between two associated components is approximately 0.5*RTT. In real networks, there may be negligible delay differences: for example, the delay between a PU and PE #1 is 5ms and the latency between the PU and PE #2 is 6ms. From the service user's perspective, such minor delay differences may be ignored and are furthermore unavoidable in Internet scenarios. Therefore, the distance parameter between two components A and B is defined as follows: Distance = DistanceStep * round( (0.5*RTT) / DistanceStep ) That is, the distance parameter is defined as the nearest integer multiple of the constant DistanceStep for the measured delay (i.e. 0.5*RTT).

In order to define a distance-aware policy, it is first necessary to define a basic rule: PEs and PUs choose "nearby" ENRP servers. Since the operation scope of RSerPool is restricted to a single organization, this condition can be met easily by appropriately locating ENRP servers. A Home ENRP server can measure the delay of the ASAP associations to its PE. As part of its ENRP updates to other ENRP servers, it can report this measured delay together with the PE information. A non-Home-ENRP server receiving such an update simply adds the delay of the ENRP association with the Home ENRP server to the PE's reported delay. Now, each ENRP server can approximate the distance to every PE in the operation scope using the equation in . Note, that delay changes are propagated to all ENRP servers upon PE re-registrations, i.e. the delay information (and the approximated distance) dynamically adapts to the state of the network.

In this section, a distance-sensitive Least Used policy is defined, based on the delay-measurement infrastructure introduced in .

The Least Used with Distance Penalty Factor (LU-DPF) policy uses load information provided by the pool elements to select the lowest-loaded pool elements within the pool. If there are multiple elements having lowest load, the nearest PE should be chosen.

The ENRP server SHOULD select at most the requested number of pool elements. Their load values SHOULD be the lowest possible ones within the pool and their distances also SHOULD be lowest. Each element MUST NOT be reported more than once to the pool user. If there is a choice of equal-loaded and equal-distanced pool elements, round robin selection SHOULD be made among these elements. The returned list of pool elements MUST be sorted by load value in ascending order (1st key) and distance in ascending order (2nd key).

The pool user should try to use the pool elements returned from the list in the order returned by the ENRP server. A subsequent call for handle resolution may result in the same list. Therefore, it is RECOMMENDED for a pool user to request multiple entries in order to have a sufficient amount of feasible backup entries available.

0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Parameter Type = 0x6 | Length = 0x14 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Policy Type = 0x40000010 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Load | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Load DPF | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Distance | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ Load: Current load of the pool element. Load DPF: The LoadDPF setting of the PE. Distance: The approximated distance in milliseconds. Between PE and Home ENRP server: The distance SHOULD be set to 0. Between Non-Home ENRP server and Home ENRP server: The delay measured on the ASAP association between Home ENRP server and PE. Between ENRP server and PU: The sums of the measured delays on the ASAP association and the ENRP association to the Home ENRP server.

The reference implementation based on the RSerPool prototype implementation rsplib can be found at .

This document does not identify security requirements beyond those already documented in the ENRP and ASAP protocol specifications.

This document does not require additional IANA actions beyond those already identified in the ENRP and ASAP protocol specifications.