@@ -14,33 +14,31 @@ See the License for the specific language governing permissions and
1414limitations under the License.
1515*/
1616
17- package queueset
18-
19- // This package implements a technique called "fair queuing for server
20- // requests". One QueueSet is a set of queues operating according to
21- // this technique.
22-
17+ // Package queueset implements a technique called "fair queuing for
18+ // server requests". One QueueSet is a set of queues operating
19+ // according to this technique.
20+ //
2321// Fair queuing for server requests is inspired by the fair queuing
2422// technique from the world of networking. You can find a good paper
2523// on that at https://dl.acm.org/citation.cfm?doid=75247.75248 or
2624// http://people.csail.mit.edu/imcgraw/links/research/pubs/networks/WFQ.pdf
2725// and there is an implementation outline in the Wikipedia article at
2826// https://en.wikipedia.org/wiki/Fair_queuing .
29-
27+ //
3028// Fair queuing for server requests differs from traditional fair
31- // queuing in three ways: (1) we are dispatching requests to be
32- // executed within a process rather than transmitting packets on a
33- // network link, (2) multiple requests can be executing at once, and
34- // (3) the service time (execution duration) is not known until the
35- // execution completes.
36-
29+ // queuing in three ways: (1) we are dispatching application layer
30+ // requests to a server rather than transmitting packets on a network
31+ // link, (2) multiple requests can be executing at once, and (3) the
32+ // service time (execution duration) is not known until the execution
33+ // completes.
34+ //
3735// The first two differences can easily be handled by straightforward
3836// adaptation of the concept called "R(t)" in the original paper and
3937// "virtual time" in the implementation outline. In that
4038// implementation outline, the notation now() is used to mean reading
4139// the virtual clock. In the original paper’s terms, "R(t)" is the
42- // number of "rounds" that have been completed at real time t, where a
43- // round consists of virtually transmitting one bit from every
40+ // number of "rounds" that have been completed at real time t ---
41+ // where a round consists of virtually transmitting one bit from every
4442// non-empty queue in the router (regardless of which queue holds the
4543// packet that is really being transmitted at the moment); in this
4644// conception, a packet is considered to be "in" its queue until the
@@ -55,12 +53,12 @@ package queueset
5553// respect to t is
5654//
5755// 1 / NEQ(t) .
58-
56+ //
5957// To generalize from transmitting one packet at a time to executing C
6058// requests at a time, that derivative becomes
6159//
6260// C / NEQ(t) .
63-
61+ //
6462// However, sometimes there are fewer than C requests available to
6563// execute. For a given queue "q", let us also write "reqs(q, t)" for
6664// the number of requests of that queue that are executing at that
@@ -79,25 +77,25 @@ package queueset
7977// real nanosecond). Where the networking implementation outline adds
8078// packet size to a virtual time, in our version this corresponds to
8179// adding a service time (i.e., duration) to virtual time.
82-
80+ //
8381// The third difference is handled by modifying the algorithm to
8482// dispatch based on an initial guess at the request’s service time
8583// (duration) and then make the corresponding adjustments once the
8684// request’s actual service time is known. This is similar, although
8785// not exactly isomorphic, to the original paper’s adjustment by
88- // `$delta` for the sake of promptness.
89-
86+ // `$\ delta$ ` for the sake of promptness.
87+ //
9088// For implementation simplicity (see below), let us use the same
9189// initial service time guess for every request; call that duration
9290// G. A good choice might be the service time limit (1
9391// minute). Different guesses will give slightly different dynamics,
9492// but any positive number can be used for G without ruining the
9593// long-term behavior.
96-
94+ //
9795// As in ordinary fair queuing, there is a bound on divergence from
9896// the ideal. In plain fair queuing the bound is one packet; in our
9997// version it is C requests.
100-
98+ //
10199// To support efficiently making the necessary adjustments once a
102100// request’s actual service time is known, the virtual finish time of
103101// a request and the last virtual finish time of a queue are not
@@ -118,3 +116,5 @@ package queueset
118116// queue’s virtual start time is advanced by G. When a request
119117// finishes being served, and the actual service time was S, the
120118// queue’s virtual start time is decremented by G - S.
119+ //
120+ package queueset
0 commit comments