64TR performance test method on cubb 25-3

Accelerated Computing Aerial Forum
1

Hi,

I tried to generate pattern 69 by using genPerfPattern.m , I’m not sure how to use the files( launch pattern files and h5 files) to do performance test on one server. Should I just use cubb_gpu_test_bench.cpp and modify parameters? The file names and content formats do not seem to meet the required specifications. The launch file includes such contents:

Cell_Configs: [TVnr_DLMIX_12232_gNB_FAPI_s0.h5]
UL_Cell_Configs: [TVnr_ULMIX_6636_gNB_FAPI_s0.h5]
config_static_harq_proc_id: 1.0
SCHED:

  • config:
    • cell_index: 0.0
      channels: [TVnr_DLMIX_12232_gNB_FAPI_s0.h5, TVnr_9364_gNB_FAPI_s0.h5]
      type: [PDSCH, PBCH, PDCCH_DL, PDCCH_UL, BFW_DL]
      slot: 0.0

Thank you!

2

@fltang You need two servers to be able to conduct cuBB level performance test. The second server runs ru-emulator and does not need a GPU. Please see here.

3

Thank you for your reply. I didn’t express myself clearly. I want to do cuphy performance test under the 64TR scenario on one server. Like Measuring cuPHY Performance using cubb_gpu_test_bench, it contains F08/F09/F14 case. Is there any way to test 64TR?

Thank you!

4

Thank you for the clarification. I thought you were aiming for cuBB performance test.

Yes, 64T64R is supported, we need to use F14 and add --rec_bf​ to enable SRS and BFW workloads; --srs_isolate​ to enable SRS run in its own MPS sub context. Since you already generated the TVs of pattern 69, here is an example of using it in this test bench. Use the attached example_perf69tv_F14.json, and run the following command:

python3 measure.py --cuphy <testBenches>/build --vectors <test_vectors> --config testcases_perf69tv_F14.json --uc uc_avg_F14_TDD.json --delay 100000 --gpu <GPU_ID> --freq <GPU_freq> --start <cell_start> --cap <cell_cap> --iterations 1 --slots <nSlots> --power <budget> --target <sms_prach> <sms_pdcch> <sms_pucch> <sms_pdsch> <sms_pusch> <sms_ssb><sms_srs>--2cb_per_sm --save_buffer --priority --prach --prach_isolate --pdcch --pdcch_isolate --pucch --pucch_isolate --tdd_pattern dddsuudddd --pusch_cascaded --ssb --ssb_isolate --csirs --groups_dl --pack_pdsch --groups_pusch --ldpc_parallel <—graph>--rec​_bf --srs_isolate

You can also use --tdd_pattern dddsuudddd_mMIMO for a different 64T64R workload timeline. All other functionalities are the same with 4T4R cases. Please note that 64T64R is experimental and the timelines are to be changed in future releases.

Thank you.
example_perf69tv_F14.json.zip (748 Bytes)

5

I will try it.

Thank you.