Just to add on top of my previous email: I believe what you're really benchmarking here is the additional overhead of running an "SFN model" partition with IPC backend, if what you're really changing in the benchmark below is just the model of the partition.

If this is the case, of course the recommendation would be just to use IPC model partitions with the IPC backendto get rid of this overhead. The recommendations made in the docs (or in the FF-M spec) where the SFN model is recommended for constrained devices is different, in the sense that is the SFN backend that is being recommended for such use cases. But if you need to use higher level of isolation, the SFN backend is not an option, and in that case you will be better using IPC model partitions with the IPC backend.

Thanks, Antonio

From: Antonio De Angelis via TF-M <tf-m@lists.trustedfirmware.org>
Sent: Friday, January 19, 2024 10:42
To: Quach, Brian <brian@ti.com>; Shebu Varghese Kuriakose <Shebu.VargheseKuriakose@arm.com>; tf-m@lists.trustedfirmware.org <tf-m@lists.trustedfirmware.org>
Cc: nd <nd@arm.com>
Subject: [TF-M] Re: SFN vs IPC partition with IPC backend
 
Hi Brian,

TF-M itself can be configured at build time to have either IPC or SFN backends. The "frontend" component mentioned in some design docs (i.e. spm_backend.rst) can't be freely chosen and is just an implementation aspect that encapsulates the mechanism to interact with the chosen backend from the same PSA APIs. In particular, if a partition is designed for the SFN model (i.e. "model": "SFN" in the partition manifest) it just means that at build time with the IPC backend there will be an additional implementation specific mechanism to allow scheduling those partitions within the IPC framework. Note that if you're using profile medium, it will automatically select the IPC backend so your profile numbers are probably just profiling the same IPC backend in both cases. Also note that a partition that supports only the IPC model, won't build at all if the selected backend is SFN (but all default partitions support the SFN model anyway). Only way to make sure that the backend is changed is to set the -DCONFIG_TFM_SPM_BACKEND to either IPC or SFN at build time.

The presentation shared in my original message shows the performance improvements that were obtained at that time between releases of TF-M with the IPC backend, and how the SFN backend compared to IPC backend in terms of cycle count on the newest (at that time) release.

Hope this helps.

Thanks, Antonio


From: Quach, Brian via TF-M <tf-m@lists.trustedfirmware.org>
Sent: Thursday, January 18, 2024 23:52
To: Quach, Brian <brian@ti.com>; Shebu Varghese Kuriakose <Shebu.VargheseKuriakose@arm.com>; tf-m@lists.trustedfirmware.org <tf-m@lists.trustedfirmware.org>
Cc: nd <nd@arm.com>; Antonio De Angelis <Antonio.DeAngelis@arm.com>
Subject: [TF-M] Re: SFN vs IPC partition with IPC backend
 

I did some of my own benchmarking with IPC backend (medium profile) and either SFN or IPC frontends.  Cycle counts measured using DWT counter.   I was expecting SFN frontend to be faster, but it’s actually slightly slower than IPC frontend at least with our platform.  Is this expected from the TFM code perspective?  

 

 

SFN frontend:

PSA API benchmarks in cycles:

psa_framework_version() call time = 340

BENCH_SP psa_version() call time = 429

BENCH_SP psa_connect() call time = 2884

BENCH_SP psa_call() set TS: call time to = 2470, from = 1155, rt = 3625

BENCH_SP psa_call() NULL 0 invecs: call time to = 1819, from = 1161, rt = 2980

BENCH_SP psa_call() NULL 1 invecs: call time to = 2025, from = 1152, rt = 3177

BENCH_SP psa_call() NULL 2 invecs: call time to = 2153, from = 1152, rt = 3305

BENCH_SP psa_call() NULL 3 invecs: call time to = 2313, from = 1155, rt = 3468

 

 

IPC frontend:
PSA API benchmarks in cycles:

psa_framework_version() call time = 339

BENCH_SP psa_version() call time = 433

BENCH_SP psa_connect() call time = 2866

BENCH_SP psa_call() set TS: call time to = 2458, from = 1147, rt = 3605

BENCH_SP psa_call() NULL 0 invecs: call time to = 1777, from = 1146, rt = 2923

BENCH_SP psa_call() NULL 1 invecs: call time to = 1974, from = 1146, rt = 3120

BENCH_SP psa_call() NULL 2 invecs: call time to = 2109, from = 1133, rt = 3242

BENCH_SP psa_call() NULL 3 invecs: call time to = 2244, from = 1136, rt = 3380

 

rt = round trip

 

Regards,
Brian

 

From: Quach, Brian via TF-M <tf-m@lists.trustedfirmware.org>
Sent: Thursday, January 18, 2024 12:08 PM
To: Shebu Varghese Kuriakose <Shebu.VargheseKuriakose@arm.com>; tf-m@lists.trustedfirmware.org
Cc: nd <nd@arm.com>; Antonio De Angelis <Antonio.DeAngelis@arm.com>
Subject: [EXTERNAL] [TF-M] Re: SFN vs IPC partition with IPC backend

 

Hi Shebu, Thanks for the info but I think you are referring to IPC vs SFN backend. IPC backend is needed for Isolation level > 1 as you said. I am using IPC backend. I’m wondering how using SFN vs IPC frontend with the IPC backend affects

ZjQcmQRYFpfptBannerStart

This Message Is From an External Sender

This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd

Hi Shebu,

 

Thanks for the info but I think you are referring to IPC vs SFN backend.  IPC backend is needed for Isolation level > 1 as you said.   I am using IPC backend.  I’m wondering how using SFN vs IPC frontend with the IPC backend affects performance & size.  All the partitions in TFM are currently SFN frontend so none of the metrics cover IPC frontend. 

 

Hi Antonio,


Thanks for the presentation but I think it only shows IPC vs SFN backends only.

 

Regards,
Brian

 

From: Shebu Varghese Kuriakose <Shebu.VargheseKuriakose@arm.com>
Sent: Thursday, January 18, 2024 7:50 AM
To: tf-m@lists.trustedfirmware.org; Quach, Brian <brian@ti.com>
Cc: nd <nd@arm.com>; Antonio De Angelis <Antonio.DeAngelis@arm.com>
Subject: [EXTERNAL] RE: SFN vs IPC partition with IPC backend

 

Hi Brian, The TF-M release notes capture the TF-M memory usage on one of the Arm reference platforms. Refer - https: //trustedfirmware-m. readthedocs. io/en/latest/releases/2. 0. 0. html It is probably worth comparing the flash and RAM usage of individual

ZjQcmQRYFpfptBannerStart

This Message Is From an External Sender

This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd

Hi Brian,

 

The TF-M release notes capture the TF-M memory usage on one of the Arm reference platforms. Refer - https://trustedfirmware-m.readthedocs.io/en/latest/releases/2.0.0.html

 

It is probably worth comparing the flash and RAM usage of individual components in ARoTless and Medium Profiles*. Both profile include almost identical services (except Protected Storage). However, ARoTless uses SFN Isolation level1 while medium profile uses IPC isolation level2.

From my understanding, SFN which is function call based is supposedly more light weight in terms of performance and memory compared to the IPC mode. Due to limited security guarantees provided by SFN, it is recommended to be used only in isolation level1. You might already know all this.

 

The PSA FF-M API calls are more expensive when in IPC mode compared to SFN model. The squad dashboard here tracks the cycle count for the various FF-M PSA API calls for IPC and SFN for various isolation levels.

https://qa-reports.linaro.org/tf/tf-m/metrics/?environment=CoreIPC&environment=Default&environment=CoreIPCTfmLevel2&environment=DefaultProfileM&environment=DefaultProfileS&metric=:summary:

 

The required build environment and API calls can be selected in the dashboard. I have selected a few for your reference - https://qa-reports.linaro.org/tf/tf-m/metrics/?environment=PERF-AN521-GCC-Level1-SFN-Release&environment=PERF-AN521-GCC-Level1-IPC-Release&environment=PERF-AN521-GCC-Level2-IPC-Release&metric=:summary:&metric=ns_psa_call&metric=ns_psa_connect&metric=ns_psa_close&metric=ns_psa_call_stateless&metric=s_psa_call&range_ns_psa_connect=96,100

 

Regards,

Shebu

 

 

* https://trustedfirmware-m.readthedocs.io/en/latest/configuration/profiles/tfm_profile_medium_arot-less.html

https://trustedfirmware-m.readthedocs.io/en/latest/configuration/profiles/tfm_profile_medium.html  

 

From: Antonio De Angelis via TF-M <tf-m@lists.trustedfirmware.org>
Sent: Thursday, January 18, 2024 9:34 AM
To: tf-m@lists.trustedfirmware.org; Quach, Brian <brian@ti.com>
Subject: [TF-M] Re: SFN vs IPC partition with IPC backend

 

Hi Brian,

 

some numbers have been shared in this presentation from a while ago:

 

Microsoft PowerPoint - TF-M Performance Improvement in v1.5.0.pptx (trustedfirmware.org)

I believe there should be more updated numbers on more recent  versions, but I don't seem to find them, maybe somebody else with more updated benchmarking can help here.

 

Thanks,

Antonio

From: Quach, Brian via TF-M <tf-m@lists.trustedfirmware.org>
Sent: Thursday, January 18, 2024 00:26
To: tf-m@lists.trustedfirmware.org <tf-m@lists.trustedfirmware.org>
Subject: [TF-M] SFN vs IPC partition with IPC backend

 

If using IPC backend, how much performance and/or memory savings is there when using SFN vs IPC partition model? 

 

I saw FF-M v1.1 recommended SFN partition model but it was not clear to me why it was preferred.

 

Regards,

 

Brian Quach

SimpleLink MCU

Texas Instruments Inc.

12500 TI Blvd, MS F-4000

Dallas, TX 75243

214-479-4076