On Wed, Mar 21, 2018 at 08:33:45PM +0200, Tal Gilboa wrote:
> Net DIM is a generic algorithm, purposed for dynamically
> optimizing network devices interrupt moderation. This
> document describes how it works and how to use it.
>
> Signed-off-by: Tal Gilboa <ta...@mellanox.com>
> ---
> Documentation/networking/net_dim.txt | 174
> +++++++++++++++++++++++++++++++++++
> 1 file changed, 174 insertions(+)
> create mode 100644 Documentation/networking/net_dim.txt
>
> diff --git a/Documentation/networking/net_dim.txt
> b/Documentation/networking/net_dim.txt
> new file mode 100644
> index 0000000..9cb31c5
> --- /dev/null
> +++ b/Documentation/networking/net_dim.txt
> @@ -0,0 +1,174 @@
> +Net DIM - Generic Network Dynamic Interrupt Moderation
> +======================================================
> +
> +Author:
> + Tal Gilboa <ta...@mellanox.com>
> +
> +
> +Contents
> +=========
> +
> +- Assumptions
> +- Introduction
> +- The Net DIM Algorithm
> +- Registering a Network Device to DIM
> +- Example
> +
> +Part 0: Assumptions
> +======================
> +
> +This document assumes the reader has basic knowledge in network drivers
> +and in general interrupt moderation.
> +
> +
> +Part I: Introduction
> +======================
> +
> +Dynamic Interrupt Moderation (DIM) (in networking) refers to changing the
> +interrupt moderation configuration of a channel in order to optimize packet
> +processing. The mechanism includes an algorithm which decides if and how to
> +change moderation parameters for a channel, usually by performing an
> analysis on
> +runtime data sampled from the system. Net DIM is such a mechanism. In each
> +iteration of the algorithm, it analyses a given sample of the data, compares
> it
> +to the previous sample and if required, it can decide to change some of the
> +interrupt moderation configuration fields. The data sample is composed of
> data
> +bandwidth, the number of packets and the number of events. The time between
> +samples is also measured. Net DIM compares the current and the previous data
> and
> +returns an adjusted interrupt moderation configuration object. In some cases,
> +the algorithm might decide not to change anything. The configuration fields
> are
> +the minimum duration (microseconds) allowed between events and the maximum
> +number of wanted packets per event. The Net DIM algorithm ascribes
> importance to
> +increase bandwidth over reducing interrupt rate.
> +
> +
> +Part II: The Net DIM Algorithm
> +===============================
> +
> +Each iteration of the Net DIM algorithm follows these steps:
> +1. Calculates new data sample.
> +2. Compares it to previous sample.
> +3. Makes a decision - suggests interrupt moderation configuration fields.
> +4. Applies a schedule work function, which applies suggested configuration.
> +
> +The first two steps are straightforward, both the new and the previous data
> are
> +supplied by the driver registered to Net DIM. The previous data is the new
> data
> +supplied to the previous iteration. The comparison step checks the difference
> +between the new and previous data and decides on the result of the last step.
> +A step would result as "better" if bandwidth increases and as "worse" if
> +bandwidth reduces. If there is no change in bandwidth, the packet rate is
> +compared in a similar fashion - increase == "better" and decrease == "worse".
> +In case there is no change in the packet rate as well, the interrupt rate is
> +compared. Here the algorithm tries to optimize for lower interrupt rate so an
> +increase in the interrupt rate is considered "worse" and a decrease is
> +considered "better". Step #2 has an optimization for avoiding false results:
> it
> +only considers a difference between samples as valid if it is greater than a
> +certain percentage. Also, since Net DIM does not measure anything by itself,
> it
> +assumes the data provided by the driver is valid.
> +
> +Step #3 decides on the suggested configuration based on the result from step
> #2
> +and the internal state of the algorithm. The states reflect the "direction"
> of
> +the algorithm: is it going left (reducing moderation), right (increasing
> +moderation) or standing still. Another optimization is that if a decision
> +to stay still is made multiple times, the interval between iterations of the
> +algorithm would increase in order to reduce calculation overhead. Also, after
I wonder if this increased interval can lead to packet drops due to
some impulse? Like, the card is receiving a low volume of packets and
suddenly a new flow starts at line rate, for example. If the max
interval is not too aggressive, this would't be a problem.
(sorry, I didn't read much of the implementation nor the drivers
already using it)
> +"parking" on one of the most left or most right decisions, the algorithm may
> +decide to verify this decision by taking a step in the other direction. This
> is
> +done in order to avoid getting stuck in a "deep sleep" scenario. Once a
> +decision is made, an interrupt moderation configuration is selected from
> +the predefined profiles.
> +
> +The last step is to notify the registered driver that it should apply the
> +suggested configuration. This is done by scheduling a work function, defined
> by
> +the Net DIM API and provided by the registered driver.
> +
> +As you can see, Net DIM itself does not actively interact with the system. It
> +would have trouble making the correct decisions if the wrong data is
> supplied to
> +it and it would be useless if the work function would not apply the suggested
> +configuration. This does, however, allow the registered driver some room for
> +manoeuvre as it may provide partial data or ignore the algorithm suggestion
> +under some conditions.
> +
> +
> +Part III: Registering a Network Device to DIM
> +==============================================
> +
> +Net DIM API exposes the main function net_dim(struct net_dim *dim,
> +struct net_dim_sample end_sample). This function is the entry point to the
> Net
> +DIM algorithm and has to be called every time the driver would like to check
> if
> +it should change interrupt moderation parameters. The driver should provide
> two
> +data structures: struct net_dim and struct net_dim_sample. Struct net_dim
> +describes the state of DIM for a specific object (RX queue, TX queue,
> +other queues, etc.). This includes the current selected profile, previous
> data
> +samples, the callback function provided by the driver and more.
> +Struct net_dim_sample describes a data sample, which will be compared to the
> +data sample stored in struct net_dim in order to decide on the algorithm's
> next
> +step. The sample should include bytes, packets and interrupts, measured by
> +the driver.
> +
> +In order to use Net DIM from a networking driver, the driver needs to call
> the
> +main net_dim() function. The recommended method is to call net_dim() on each
> +interrupt. Since Net DIM has a built-in moderation and it might decide to
> skip
> +iterations under certain conditions, there is no need to moderate the
> net_dim()
> +calls as well. As mentioned above, the driver needs to provide an object of
> type
> +struct net_dim to the net_dim() function call. It is advised for each entity
> +using Net DIM to hold a struct net_dim as part of its data structure and use
> it
> +as the main Net DIM API object. The struct net_dim_sample should hold the
> latest
> +bytes, packets and interrupts count. No need to perform any calculations,
> just
> +include the raw data.
> +
> +The net_dim() call itself does not return anything. Instead Net DIM relies on
> +the driver to provide a callback function, which is called when the algorithm
> +decides to make a change in the interrupt moderation parameters. This
> callback
> +will be scheduled and run in a separate thread in order not to add overhead
> to
> +the data flow. After the work is done, Net DIM algorithm needs to be set to
> +the proper state in order to move to the next iteration.
> +
> +
> +Part IV: Example
> +=================
> +
> +The following code demonstrates how to register a driver to Net DIM. The
> actual
> +usage is not complete but it should make the outline of the usage clear.
> +
> +my_driver.c:
> +
> +#include <linux/net_dim.h>
> +
> +/* Callback for net DIM to schedule on a decision to change moderation */
> +void my_driver_do_dim_work(struct work_struct *work)
> +{
> + /* Get struct net_dim from struct work_struct */
> + struct net_dim *dim = container_of(work, struct net_dim,
> + work);
> + /* Do interrupt moderation related stuff */
> + ...
> +
> + /* Signal net DIM work is done and it should move to next iteration */
> + dim->state = NET_DIM_START_MEASURE;
Doesn't this need some sort of order guarantee? This assignment cannot
happen before the complete usage of 'dim' struct, otherwise the
interrupt handler may overwrite stuff that is still being read.
Or maybe this 2-steps to measure handles it? (START_MEASURE ->
MEASURE_IN_PROGRESS) As it will only overwrite something in the next
irq, it wouldn't be able to race the workqueue. As work-queues can be
preempted, this doesn't seem enough.
> +}
> +
> +/* My driver's interrupt handler */
> +int my_driver_handle_interrupt(struct my_driver_entity *my_entity, ...)
> +{
> + ...
> + /* A struct to hold current measured data */
> + struct net_dim_sample dim_sample;
> + ...
> + /* Initiate data sample struct with current data */
> + net_dim_sample(my_entity->events,
> + my_entity->packets,
> + my_entity->bytes,
> + &dim_sample);
> + /* Call net DIM */
> + net_dim(&my_entity->dim, dim_sample);
> + ...
> +}
> +
> +/* My entity's initialization function (my_entity was already allocated) */
> +int my_driver_init_my_entity(struct my_driver_entity *my_entity, ...)
> +{
> + ...
> + /* Initiate struct work_struct with my driver's callback function */
> + INIT_WORK(&my_entity->dim.work, my_driver_do_dim_work);
> + ...
> +}
> --
> 1.8.3.1
>
