fan tuning

how i got truenas1 quiet enough to leave its closet door open, and how claude tested it. the closet opens onto the home theatre, and it is not climate controlled.

where it started

the board's BMC ran every fan from FSC_INDEX, one index it computes for the whole box, through one curve. so every fan answered every heat source. the CPU fan ran near 6000 RPM at idle, and nothing sped up for the GPU in particular.

where it ended

each fan follows the sensor that matches what it cools, through its own curve:

header	fan	follows	idle before	idle now
FAN1	CPU cooler	TEMP_CPU, table 5	5800-6400 RPM	2800 RPM
FAN2, FAN3	drive wall	TEMP_OUTLET2, table 6	1200-1400 RPM	1000 RPM
FAN4	60 mm fan over the PCIe slots	nothing, fixed at 100%	3000-3200 RPM	3200 RPM
FAN5	rear	TEMP_OUTLET2, table 6	1000 RPM	800 RPM
FAN6	the M.2 card's own fan	TEMP_OUTLET2, table 7	2200-2400 RPM	2400 RPM

• air runs front to back. the drive wall pulls it over the drives and pushes it across the whole board side, past the CPU cooler, the memory, the GPU and the PCIe cards, and the rear fan takes it out. so the drive wall is the main airflow for everything on the board, not just the drives
• the board sits turned in the case, a Sliger CX4712, so its headers run FAN6 to FAN1 from left to right
• the slot fan stays at full speed. NVMe drives on PCIe adapters cook without airflow, and the boot drives' controllers sit near 80 °C even with it

the BMC's temperature sensors, with the closet at about 29 °C:

sensor	idle	GPU at its 350 W cap	fans on it
TEMP_CPU	58-60 °C	62-64 °C	FAN1
TEMP_OUTLET2	38-39 °C	49-50 °C	FAN2, FAN3, FAN5, FAN6
TEMP_OUTLET1	48-49 °C	48-49 °C	nothing
FSC_INDEX	63-65 °C	67-69 °C	the factory profile
TEMP_MB2	36-37 °C	49-50 °C	nothing
TEMP_MB1	34 °C	36 °C	nothing
TEMP_LAN	64-66 °C	64-67 °C	nothing
TEMP_M.2_1, TEMP_M.2_2	46-48 °C	54-56 °C	nothing
TEMP_DDR5_*, 8 DIMMs	40-51 °C	43-50 °C	nothing

• TEMP_OUTLET2 is the one that moves with the GPU, so the case fans and the M.2 card's fan follow it. TEMP_OUTLET1 hardly moves whatever the load
• the BMC sees none of the add-in NVMe drives, the hard drives, the GPU or the network card. glances shows those, see glances

the curves, as °C duty%:

table	for	steps	hysteresis
5	CPU	25 20, 66 25, 70 30, 74 40, 78 55, 82 70, 85 85, 88 100	5 °C
6	case fans	25 35, 42 40, 46 50, 50 65, 53 80, 56 100	3 °C
7	M.2 card	25 50, 42 60, 45 70, 48 85, 51 100	3 °C

• each curve is a customized open-loop table, tied to its sensor and fans by a customized assignment, and each of those fans is set to customized. a fan can be in only one assignment, so take it out of the BMC's own first
• the CPU curve first rises at 66 °C, above where the CPU sits at idle and at half load, with 5 °C of hysteresis and a gradual ramp up, so the fan holds a speed rather than hunting between two
• at idle the case fans run at 35% and the M.2 card's fan at 50%, which i cannot hear. they speed up only when the GPU heats the case
• each table starts at 25 °C, below any reading the box gives, so nothing rests on what the BMC does below a table's first step
• redfish reports the fans but cannot set them. the BMC's web pages, and the API behind them, can
• a BMC reset or firmware update can bring the factory profile back, so the settings live in a file in my repo, with a script that checks the BMC against it and restores it. gatus watches the slot fan and the card's fan over redfish

what it does under load

load	result
a sustained write to fast	the Seagates level off at 60 °C; they are rated to 70 °C
a sustained mixed load on rust	the IronWolfs rise 1 to 2 °C
half the CPU's threads busy	the CPU holds 61 to 65 °C
the GPU at its 350 W cap	the GPU levels off at 81 °C, and the closet rises 2 °C
a sustained write to fast with the GPU busy	the Seagates pass 70 °C in about 7 minutes, even with the card's fan at full speed, because the GPU heats the air the card draws

• nothing here loads fast that hard. the heaviest jobs, PBS's verifies and the offsite copy, read rust

how claude tested it

the approach, not the scripts:

• before any change it saved the BMC's fan settings. a guard put them back the moment anything reached its limit
• a logger recorded every component once a minute: the fans and the BMC's sensors over redfish with a read-only account, every drive and card from glances, the GPU from nvidia-smi, and the closet from the UPS card's environment sensor
• each limit sat at or below the part's rating: 70 °C for the Seagates, 45 °C for the hard drives, 85 °C for the CPU and the GPU, 35 °C for the closet, and any fan under 300 RPM
• it changed the fans through the BMC's web API and read every change back. for the M.2 card's fan i stood at the server and listened at each step
• it loaded the box one source at a time, 12 minutes each with a cool-down between: a write to each pool on a throwaway dataset, half the CPU's threads, and the local sysadmin model writing runbooks on the GPU. then all of them at once. a watchdog stopped every load at the first limit, which is how the last row above ends
• the first pass showed the CPU fan hunting between two speeds and the case fans deaf to the GPU, so it changed the curves and ran the GPU and combined loads again