Matter RAM Requirements

[doru91]

On NXP K32W061 (Thread + BLE) I reached a high level of RAM consumption, around 141.5K (Thread FTD). This is distributed as follows:

.heap 61440 bss 75240 .data 1796

HEAP

• FreeRtos tasks's stacks (which are dynamically allocated) requires around 19.5K of RAM (see the picture below). Fine tuning of the stacks was done using FreeRtos Stack High Watermark feature some time ago.

• The main problem is that immediately after I start the device (no BLE/Thread communication available yet) another 30K (!) of RAM are allocated (see the picture below).

• At a first glance, the main consumer is the InitServer() function, specifically the initialization of each fabrics requires around ~ 1.4K and there are 16 Fabrics (CHIP_CONFIG_MAX_DEVICE_ADMINS). I guess we don't need 16 Fabrics/device, so I believe this number can be lowered to a more realistic value like 4 Fabrics/device (and we save around 17K of RAM - I see that Qorvo already operated this change). Even with this number:

  1. how many CASE session/device are expected and how much RAM does a CASE session expects?
  2. how many ACL lists do we need to store per device and how much RAM requires a single ACL list?

BSS

here are the main consumers:

SECTION: .bss SECTION: .bss 100% 75240 total ├── 18% 13280 _ZN2ot12gInstanceRawE ├── 14% 10371 memp_memory_PBUF_POOL_base ├── 6% 4620 rwip_heap_non_ret ├── 5% 4096 au8PDM_InternalFlashDataBuffer ├── 5% 4096 au8PDM_SegmentDataBuffer ├── 4% 3080 _ZN4chip3app23sInteractionModelEngineE ├── 4% 2960 _ZN12_GLOBAL__N_18gFabricsE ├── 3% 2128 memHeap ├── 3% 2120 _ZN12_GLOBAL__N_19gSessionsE

• If the device is downgraded to SED (from FTD) then .bss decreases with around 5K;

• OPENTHREAD_CONFIG_NUM_MESSAGE_BUFFERS is set to the default 44. Each 10 buffers requires ~1.3K of RAM. I believe it really depends on the Thread device used. For an SED, I see that some vendors choose to set this variable to 10 (so _ZN2ot12gInstanceRawE was decreased with around 4K):

1. Is there a minimum recommended size for the numbers of buffers if the device is an SED?
2. Same question if the device is an FTD.

• if LWIP RAW, LWIP TCPEndpoints are disabled and PBUF_POOL_SIZE is decreased from 8 to 6 then another 3K of .bss can be saved. I see that some vendors operated this change.
  1. Do we need TCP for Matter?
  2. What is the recommended LWIP Pool Size for Matter?

In the end, I believe that the current minimal RAM requirements are pretty high for a Matter device and we need to understand better where each byte of memory is consumed.

[tcarmelveilleux]

• Regarding large amount of per-fabric RAM data:

  • This can mostly go away if we use "just in time" reading of certs (e.g. as proposed in #7695). I know @pan-apple is considering this. It would save 400x3 bytes (3 certs, 1200 bytes) of RAM per fabric since those certs are only needed during CASE Sigma2/3 processing, and can be loaded 1 by 1 as needed.

• Regarding the minimum number of fabrics:

  • The minimum number of fabrics is likely 4-5 to support. The consensus is not done, but 16 is too much for a RAM constrained device. See https://github.com/CHIP-Specifications/connectedhomeip-spec/issues/2551.

• Regarding number of CASE session and ACL entries, this is still being discussed (https://github.com/CHIP-Specifications/connectedhomeip-spec/issues/2551) but at the same time, ACL entries do not necessarily all need to be in RAM all the time if there is an efficient way to cache them in flash to read them "in-place". This would be feasible if the implementation of ACL entry lookup is abstracted in a way that it can be optimized per platform/product.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.