Description
Hello, I ran the nvm-block-bench and there is no response after the message “at launch kernel”.
~/bam/build$ sudo ./bin/nvm-block-bench --threads=1 --blk_size=2 --reqs=1 --pages=2048 --queue_depth=2 --page_size=512 --num_blks=16384 --gpu=0 --n_ctrls=1 --num_queues=1 --random=false
SQs: 128 CQs: 128 n_qps: 1
n_ranges_bits: 6
n_ranges_mask: 63
pages_dma: 0x7f923b270000 4000a70000
HEREN
Cond1
800 8 8 100
Finish Making Page Cache
finished creating cache
0000:01:00.0
atlaunch kernel
summary of the problem:
・After “atlaunch kernel,” the application appears to
hang or never complete.
・We have tried reducing queue depth to 2 (SQ=2,
CQ=2) and limiting the number of commands to 1
(--reqs=1), but the kernel still does not proceed.
・Debug logs show that the queue is created, but no
NVMe completion entries ever arrive.
・The program remains stuck in a loop waiting for a
locked queue entry to be unlocked.
・We suspect that the NVMe commands are not
actually reaching the controller or the completion is
never written back.
test environment:
Linux 5.15.0-125-generic
CPU: 12th Gen Intel(R) Core(TM) i7-12700
GPU: NVIDIA RTX 4000 Ada Generation
I know Tesla/Datacenter grade GPU is needed and RTX 4000 Ada won’t work as it is due to lack of BAR space, but please teach me what is underlying problem when using RTX 4000 Ada or give me some advices.