1 Net DIM - Generic Network Dynamic Interrupt Moderation
2 ======================================================
5 Tal Gilboa <talgi@mellanox.com>
13 - The Net DIM Algorithm
14 - Registering a Network Device to DIM
18 ======================
20 This document assumes the reader has basic knowledge in network drivers
21 and in general interrupt moderation.
25 ======================
27 Dynamic Interrupt Moderation (DIM) (in networking) refers to changing the
28 interrupt moderation configuration of a channel in order to optimize packet
29 processing. The mechanism includes an algorithm which decides if and how to
30 change moderation parameters for a channel, usually by performing an analysis on
31 runtime data sampled from the system. Net DIM is such a mechanism. In each
32 iteration of the algorithm, it analyses a given sample of the data, compares it
33 to the previous sample and if required, it can decide to change some of the
34 interrupt moderation configuration fields. The data sample is composed of data
35 bandwidth, the number of packets and the number of events. The time between
36 samples is also measured. Net DIM compares the current and the previous data and
37 returns an adjusted interrupt moderation configuration object. In some cases,
38 the algorithm might decide not to change anything. The configuration fields are
39 the minimum duration (microseconds) allowed between events and the maximum
40 number of wanted packets per event. The Net DIM algorithm ascribes importance to
41 increase bandwidth over reducing interrupt rate.
44 Part II: The Net DIM Algorithm
45 ===============================
47 Each iteration of the Net DIM algorithm follows these steps:
48 1. Calculates new data sample.
49 2. Compares it to previous sample.
50 3. Makes a decision - suggests interrupt moderation configuration fields.
51 4. Applies a schedule work function, which applies suggested configuration.
53 The first two steps are straightforward, both the new and the previous data are
54 supplied by the driver registered to Net DIM. The previous data is the new data
55 supplied to the previous iteration. The comparison step checks the difference
56 between the new and previous data and decides on the result of the last step.
57 A step would result as "better" if bandwidth increases and as "worse" if
58 bandwidth reduces. If there is no change in bandwidth, the packet rate is
59 compared in a similar fashion - increase == "better" and decrease == "worse".
60 In case there is no change in the packet rate as well, the interrupt rate is
61 compared. Here the algorithm tries to optimize for lower interrupt rate so an
62 increase in the interrupt rate is considered "worse" and a decrease is
63 considered "better". Step #2 has an optimization for avoiding false results: it
64 only considers a difference between samples as valid if it is greater than a
65 certain percentage. Also, since Net DIM does not measure anything by itself, it
66 assumes the data provided by the driver is valid.
68 Step #3 decides on the suggested configuration based on the result from step #2
69 and the internal state of the algorithm. The states reflect the "direction" of
70 the algorithm: is it going left (reducing moderation), right (increasing
71 moderation) or standing still. Another optimization is that if a decision
72 to stay still is made multiple times, the interval between iterations of the
73 algorithm would increase in order to reduce calculation overhead. Also, after
74 "parking" on one of the most left or most right decisions, the algorithm may
75 decide to verify this decision by taking a step in the other direction. This is
76 done in order to avoid getting stuck in a "deep sleep" scenario. Once a
77 decision is made, an interrupt moderation configuration is selected from
78 the predefined profiles.
80 The last step is to notify the registered driver that it should apply the
81 suggested configuration. This is done by scheduling a work function, defined by
82 the Net DIM API and provided by the registered driver.
84 As you can see, Net DIM itself does not actively interact with the system. It
85 would have trouble making the correct decisions if the wrong data is supplied to
86 it and it would be useless if the work function would not apply the suggested
87 configuration. This does, however, allow the registered driver some room for
88 manoeuvre as it may provide partial data or ignore the algorithm suggestion
89 under some conditions.
92 Part III: Registering a Network Device to DIM
93 ==============================================
95 Net DIM API exposes the main function net_dim(struct dim *dim,
96 struct dim_sample end_sample). This function is the entry point to the Net
97 DIM algorithm and has to be called every time the driver would like to check if
98 it should change interrupt moderation parameters. The driver should provide two
99 data structures: struct dim and struct dim_sample. Struct dim
100 describes the state of DIM for a specific object (RX queue, TX queue,
101 other queues, etc.). This includes the current selected profile, previous data
102 samples, the callback function provided by the driver and more.
103 Struct dim_sample describes a data sample, which will be compared to the
104 data sample stored in struct dim in order to decide on the algorithm's next
105 step. The sample should include bytes, packets and interrupts, measured by
108 In order to use Net DIM from a networking driver, the driver needs to call the
109 main net_dim() function. The recommended method is to call net_dim() on each
110 interrupt. Since Net DIM has a built-in moderation and it might decide to skip
111 iterations under certain conditions, there is no need to moderate the net_dim()
112 calls as well. As mentioned above, the driver needs to provide an object of type
113 struct dim to the net_dim() function call. It is advised for each entity
114 using Net DIM to hold a struct dim as part of its data structure and use it
115 as the main Net DIM API object. The struct dim_sample should hold the latest
116 bytes, packets and interrupts count. No need to perform any calculations, just
117 include the raw data.
119 The net_dim() call itself does not return anything. Instead Net DIM relies on
120 the driver to provide a callback function, which is called when the algorithm
121 decides to make a change in the interrupt moderation parameters. This callback
122 will be scheduled and run in a separate thread in order not to add overhead to
123 the data flow. After the work is done, Net DIM algorithm needs to be set to
124 the proper state in order to move to the next iteration.
130 The following code demonstrates how to register a driver to Net DIM. The actual
131 usage is not complete but it should make the outline of the usage clear.
135 #include <linux/dim.h>
137 /* Callback for net DIM to schedule on a decision to change moderation */
138 void my_driver_do_dim_work(struct work_struct *work)
140 /* Get struct dim from struct work_struct */
141 struct dim *dim = container_of(work, struct dim,
143 /* Do interrupt moderation related stuff */
146 /* Signal net DIM work is done and it should move to next iteration */
147 dim->state = DIM_START_MEASURE;
150 /* My driver's interrupt handler */
151 int my_driver_handle_interrupt(struct my_driver_entity *my_entity, ...)
154 /* A struct to hold current measured data */
155 struct dim_sample dim_sample;
157 /* Initiate data sample struct with current data */
158 dim_update_sample(my_entity->events,
163 net_dim(&my_entity->dim, dim_sample);
167 /* My entity's initialization function (my_entity was already allocated) */
168 int my_driver_init_my_entity(struct my_driver_entity *my_entity, ...)
171 /* Initiate struct work_struct with my driver's callback function */
172 INIT_WORK(&my_entity->dim.work, my_driver_do_dim_work);