Storage performance benchmarking was done by using the FIO linux tool which ensures data randomness, avoids OS level caching in order to benchmark only the underlying storage. All the tests were run separately to ensure clarity of results and benchmark certain processes in isolation.
fio --name=fiotest --filename=/nvme/small_fio_test_read --size=16Gb --rw=read --bs=4K --direct=1 --numjobs=1 --ioengine=libaio --iodepth=1 --group_reporting --runtime=60 --startdelay=60
The filename should be on path /nvme/ in order for the filename to be created on SSD and the proper bandwidth to be tested. To see the storage devices on your device and their paths utilize:
lsblk -o KNAME,TYPE,SIZE,MODEL
In our experiments we were mainly changing the block size (bs) and number of I/O units to keep in flight against the file (iodepth).
Considering the crucial role of efficient models and data loading/unloading for these vehicles' operational performance, we hypothesized that the bandwidth utilization between SSDs and DRAM directly affects the speed of these processes. Specifically, we proposed that higher bandwidth occupancy would correlate with quicker data handling speeds, particularly for loading and offloading models.
The hypothesis in this experiment was that increasing the load on the PCIe bus would enhance bandwidth utilization, maximizing the bus's capacity and accelerating data transfer rates. This was confirmed as the experiments in figure below showed a significant increase in bandwidth with larger block sizes. Notably, the bandwidth stabilized around the maximum theoretical rate of approximately 4 GB/s, per the specifications of the PCIe bus on the Jetson AGX Orin platform, upon reaching a block size of 0.5 GB.
.png)
Analysis of smaller block sizes (below 15KB) in figure below revealed a gradual increase in bandwidth, illustrating poor occupation of the PCIe bus with minimal I/O load; for instance, a block size of 64KB achieved only 0.09 GB/s. These findings confirm that to fully utilize the PCIe's capacity on the Jetson AGX Orin, a minimum block size of 0.5 GB is necessary, and it must be processed concurrently to achieve optimal data transfer rates. This experiment substantiates our hypothesis that larger block sizes result in more efficient utilization of the available bandwidth on the PCIe bus.
In our experimental setup, we aimed to test the read speed of eMMC flash memory, commonly used in AV pipelines where rapid data transfer to DRAM for GPU processing is crucial. Using the Unix 'dd' command, we conducted tests transferring various data sizes with different block sizes (bs param). This approach tested the hypothesis that larger data chunks transferred simultaneously could maximize bandwidth utilization and accelerate the transfer process. The specific paths to the files on the eMMC flash memory were identified using the lsblk -o KNAME,TYPE,SIZE,MODEL command (as before in SSD analysis). For eMMC Flash Memory the storage device's path starts with mmcblk.
dd if=/dev/mmcblk3 of=/dev/null bs=1G count=1 iflag=direct
bs - block size
count - number of blocks
iflag - parameter to set up the i/o mode
Our experimental results strongly support the hypothesis that smaller block sizes only partially utilize the bandwidth between eMMC flash memory and DRAM, resulting in suboptimal performance. This is particularly evident when using a 1KB block size, where bandwidth usage could be much better (the average bandwidth, according to the table, is 4 MB/s). Conversely, a block size of 4096 KB (4 MB) and bigger maximize bandwidth utilization, achieving speeds around 270 MB/s. Importantly, increasing the block size further provides no substantial improvement in bandwidth utilization. For instance, block sizes larger than 64 MB (64000 KB) yield only some of the insights and tend to clutter the graphical representation of our data. Therefore, our data unequivocally demonstrates that transferring larger data chunks—greater than 4 MB—optimizes bandwidth usage, thereby enhancing data transfer efficiency between eMMC and DRAM.
