commit | 67efa244374f22c9c3db1adacf97919a11335874 | [log] [tgz] |
---|---|---|
author | Patrick Venture <venture@google.com> | Thu Oct 11 09:44:42 2018 -0700 |
committer | Patrick Venture <venture@google.com> | Thu Oct 11 09:44:42 2018 -0700 |
tree | e4bc7e6c752ce5d2dc27f0c02938a4c260850e4f | |
parent | cc74ae198ac3b48013cd470edf5f409fd1463267 [diff] |
docs: Cleanup of Design Doc The google-drive -> markdown added those reference points and they are not needed and show up. This removes them. Change-Id: If7ab88a08b048e78b4cd8a2159e7e56180bd55a3 Signed-off-by: Patrick Venture <venture@google.com>
Develop a tray level fan control system that will use exhaust temperature and other machine temperature information to control fan speeds in order to keep machines within acceptable operating conditions.
Effectively porting the Chromium EC thermal code to run on the BMC and use the OpenBMC dbus namespace and IPMI commands.
Recent server systems come with a general secondary processing system attached for the purpose of monitoring and control, generally referred to as a BMC[^2]. There is a large effort to develop an open source framework for writing applications and control systems that will run on the BMC, known as OpenBMC[^3],[^4]. Within Google the effort has been internalized (while also providing upstream pushes) as gBMC[^5]. The primary goal of OpenBMC is to provide support for remote and local system management through a REST interface, and also through IPMI[^6] tied together via the system dbus. OpenBMC provides many applications and daemons that we can leverage and improve.
The BMC is wired such that it has direct access and control over many motherboard components, including fans and temperature sensors[^7]. Therefore, it is an ideal location to run a thermal control loop, similar to the EC. However, to upstream it will need to follow (as best as possible) the OpenBMC specifications for communicating and executing[^8].
IPMI allows for OEM commands to provide custom information flow or system control with a BMC. OEM commands are already lined up for certain other accesses routed through the BMC, and can be upstreamed for others to use.
The BMC will run a daemon that controls the fans by pre-defined zones. The application will use thermal control, such that each defined zone is kept within a range and adjusted based on thermal information provided from locally readable sensors as well as host-provided information over an IPMI OEM command.
A system (or tray) will be broken out into one or more zones, specified via configuration files or dbus. Each zone will contain at least one fan and at least one temperature sensor and some device margins. The sensor data can be provided via sysfs, dbus, or through IPMI. In either case, default margins should be provided in case of failure or other unknown situation.
The system will run a control loop for each zone with the attempt to maintain the temperature within that zone within the margin for the devices specified.
The software will run as a multi-threaded daemon that runs a control loop for each zone, and has a master thread which listens for dbus messages. Each zone will require at least one fan that it exclusively controls, however, zones can share temperature sensors.
In this figure the communications channels between swampd and ipmid and phosphor-hwmon are laid out.
To be upstreamed to OpenBMC for use on open-power systems, we need to follow the OpenBMC code style specification[^9] and leverage the dbus framework for reading sensors and fan control[^10].
There is already a daemon, which given a configuration file for a hwmon device, will add it to the dbus objects namespace which handles queries for values such a temperature or fan speed and allows another process to control the fan speed[^11]. It is the goal to utilize this other daemon through dbus to read the onboard sensors and control the fans.
Because of the present implementation of the dbus interfaces to require controlling the fans only via specifying the RPM target, whereas the driver we're using for Quanta-Q71l (the first system) only allows writing PWM. This can be controlled either directly or via dbus.
A configuration file will need to exist for each board, likely in YAML[^12]. Similar information will also be necessary for gsys, such that it knows what sensors to read and send to the BMC. Presently it does something similar with EC, so it shouldn't be unreasonable to do something similar.
Each zone must have at least one fan that it exclusively controls. Each zone must have at least one temperature sensor, but they may be shared.
The external devices specified in the zone must have default information as a fallback, while their current temperatures will be provided by gsys. Some devices adapt quickly and others slowly, and this distinction will need to be a factor and described in the configuration.
The internal thermometers specified will be read via sysfs.
{ZONEID}: {PIDID}: type: "fan" | "margin" ipmi: {IPMI_ID} name: "fan1" readpath: "/xyz/openbmc_project/sensors/fan_tach/fan1" writepath: "/sys/class/hwmon/hwmon0/pwm0" pidinfo: samplerate: 0.1 // sample time in seconds p_coeff: 0.01 // coefficient for proportional i_coeff: 0.001 // coefficient for integral integral_limit: min: 0 max: 100 output_limit: min: 0 max: 100 slew_negative: 0 slew_positive: 0 {PIDID}: type: "margin" ipmi: {IPMI_ID} name: "sluggish0" readpath: "/xyz/openbmc_project/sensors/external/sluggish0" writepath: "" pidinfo: samplerate: 1 // sample time in seconds p_coeff: 94.0 i_coeff: 2.0 integral_limit: min: 3000 max: 10000 output_limit: min: 3000 max: 10000 slew_negative: 0 slew_positive: 0
Due to data center requirements, the delta between the outgoing air temperature and the environmental air temperature must be no greater than 15C.
Gsys needs the ability to send to the BMC, the margin information on the devices that it knows how to read that the BMC cannot. There is no command in IPMI that currently supports this use-case, therefore it will be added as an OEM command.
The state of the BMC readable temperature sensors can be read through normal IPMI commands and is already supported.
Gsys needs to be able to set the control of the thermal system to either automatic or manual. When manual, the daemon will effectively wait to be told to be put back in automatic mode. It is expected in this manual mode that something will be controlling the fans via the other commands.
Manual mode is controlled by zone through the following OEM command:
Byte | Purpose | Value |
---|---|---|
00 | netfn | 0x2e |
01 | command | 0x04 (also using manual command) |
02 | oem1 | 0xcf |
03 | oem2 | 0xc2 |
04 | padding | 0x00 |
05 | SubCommand | Get or Set. Get == 0, Set == 1 |
06 | ZoneId | |
07 | Mode | If Set, Value 1 == Manual Mode, 0 == Automatic Mode |
Byte | Purpose | Value |
---|---|---|
02 | oem1 | 0xcf |
03 | oem2 | 0xc2 |
04 | padding | 0x00 |
07 | Mode | If Set, Value 1 == Manual Mode, 0 == Automatic Mode |
Gbmctool needs to be able to read back whether a zone is in failsafe mode. This setting is read-only because it's dynamically determined within Swampd per zone.
Byte | Purpose | Value |
---|---|---|
00 | netfn | 0x2e |
01 | command | 0x04 (also using manual command) |
02 | oem1 | 0xcf |
03 | oem2 | 0xc2 |
04 | padding | 0x00 |
05 | SubCommand | Get == 2 |
06 | ZoneId |
Byte | Purpose | Value |
---|---|---|
02 | oem1 | 0xcf |
03 | oem2 | 0xc2 |
04 | padding | 0x00 |
07 | failsafe | 1 == in Failsafe Mode, 0 not in failsafe mode |
Gsys needs to update the thermal controller with information not necessarily available to the BMC. This will comprise of a list of temperature (or margin?) sensors that are updated by the set sensor command. Because they don't represent real sensors in the system, the set sensor handler can simply broadcast the update as a properties update on dbus when it receives the command over IPMI.
Gsys can override a specific fan's PWM when we implement the set sensor IPMI command pathway.
Gsys can read fan_tach through the normal IPMI interface presently exported for sensors.
The plan is to listen for fan_tach updates for each fan in a background thread. This will receive an update from phosphor-hwmon each time it updates any sensor it cares about.
By default phosphor-hwmon reads each sensor in turn and then sleeps for 1 second. We'll be updating phosphor-hwmon to sleep for a shorter period -- how short though is still TBD. We'll also be updating phosphor-hwmon to support pwm as a target.
Each zone will require a control loop that monitors the associated thermals and controls the fan(s). The EC PID loop is designed to hit the fans 10 times per second to drive them to the desired value and read the sensors once per second. We'll be receiving sensor updates with such regularly, however, at present it takes ~0.13s to read all 8 fans. Which can't be read constantly without bringing the system to its knees -- in that all CPU cycles would be spent reading the fans. TBD on how frequently we'll be reading the fan sensors and the impact this will have.
The main thread will manage the other threads, and process the initial configuration files. It will also register a dbus handler for the OEM message.
By default, swampd isn't compiled to log information. To compile it for tuning, you'll need to add:
EXTRA_OEMAKE_append_YOUR_MACHINE = " CXXFLAGS='${CXXFLAGS} -D__TUNING_LOGGING__'"
To the recipe.
This project is designed to be a daemon running within the OpenBMC environment. It will use a well-defined configuration file to control the temperature of the tray components to keep them within operating conditions. It will require coordinate with gsys and OpenBMC. Providing a host-side service upstream to talk to the BMC is beyond the scope of this project.
A rogue client on the host could send invalid thermal information causing physical damage to the system. There will be an effort to sanity check all input from gsys to alleviate this concern.
This device holds no user data, however, you could profile the types of jobs executed on the server by watching its temperatures.
Testing individual code logic will be handled through unit-tests, however some pieces of code rely on abstractions such that we can swap out dbus with something encapsulated such that testing can be done without strictly running on a real system.
Testing the system on real hardware will be performed to verify:
Unit-tests will provide that we know it validates information from gsys properly as well as handles difficult to reproduce edge cases.
The testing of this project on real hardware can likely fold into the general gBMC testing planned.
[^2]: BMC - Board Management Controller [^3]: with url https://github.com/openbmc/openbmc [^4]: with url https://github.com/facebook/openbmc [^5]: with url http://go/gbmc [^6]: with url http://www.intel.com/content/www/us/en/servers/ipmi/ipmi-second-gen-interface-spec-v2-rev1-1.html [^7]: Excluding temperature sensors on PCIe cards and other add-ons. [^8]: They prefer c++. [^9]: With url https://github.com/openbmc/docs/blob/master/cpp-style-and-conventions.md [^10]: with url https://github.com/openbmc/phosphor-dbus-interfaces [^11]: with url https://github.com/openbmc/phosphor-hwmon [^12]: YAML appears to be the configuration language of choice for OpenBMC.