Docs

Torch Object Storage

TOS accelerator

Measuring accelerator performance

This topic describes the observed performance differences when using TOS's intranet domain name versus the accelerated domain name in specific business scenarios.

Bulk data operations

Test method

We tested the read performance using tosutil in two scenarios: accessing TOS via its intranet and accelerator domain names. The test dataset consisted of 10,000 100-KB objects, totaling approximately 976 MB.

We used the tosutil cp command to download this set of 100-KB objects from TOS to a local directory for both the intranet and the accelerated domain names. Then, we compared the performance metrics between the two scenarios.

tosutil parameters: The default parameters were used.

TOS intranet domain name: ./tosutil cp tos://accelerator-tos-rd-test-beijing-azd/100k_files ./temp -r.

TOS accelerated domain name: ./tosutil cp tos://accelerator-tos-rd-te-01987847e6127cc48d150322293dafc8-tosalias/100k_files ./temp -r.

Test results

Item	Average latency (ms)	QPS	Bandwidth (MB/s)
TOS intranet domain name	7.53	372	36.60
TOS accelerated domain name	3.17	856	85.14

Using tosutil alongside a TOS accelerator improved performance by about 2.3 times compared to not using the accelerator. The TOS accelerator can boost data transfer and access speeds for bulk data operations.

Big data (HNS)

Test method

We performed nnbench and DFSIO benchmark operations against both the intranet and the accelerated domain names using Hadoop with Proton. Then, we compared the key read metrics of the two scenarios.

nnbench test results

We used 100 maps, each accessing 5,000 objects, for a total of 500,000 1-MB objects.

Item	Read latency	QPS
TOS intranet domain name	65 ms	1,280
TOS accelerated domain name	10 ms	4,982

tip

Read latency refers to the delay that occurs during a GetObject operation.

The TOS accelerator increased QPS by nearly fourfold and reduced read latency to about one-sixth of the value without the accelerator.

DFSIO test results

We performed read operations on a test dataset consisting of 100 objects, each 100 MB in size.

Item	Throughput
TOS intranet domain name	99.86 MB/s
TOS accelerated domain name	252.24 MB/s

Machine learning

Test method

We simulated the loading of machine learning data using the TOS Torch Connector in order to compare TOS's read performance via the intranet and the accelerated domain names. The test dataset consisted of two million images, averaging 120 KB each, for a total of approximately 228 GB.

Performance scales linearly in multi-machine, multi-process deployments. Here, a single-process setup is used as an example.

Test parameters:

batch_size: 256

num_workers: 32

tostorchconnector==1.1.2

Test results

Item	Average batch load latency (ms)	QPS (img/s)	Bandwidth (MB/s)
TOS intranet domain name	160.07	1,600	192
TOS accelerated domain name	32.12	7,936	952

The TOS accelerator improved performance by about five times compared to not using it. The results indicate that, in comparable AI scenarios, leveraging the TOS accelerator can lead to enhanced throughput and reduced latency.

Measuring response latency during downloads

Test method

To test latency, we downloaded an object of about 10 MB using both the intranet and the accelerated domain names, and we recorded the response latency percentiles.

Test results

Item	Average latency (ms)	P50	P90	P95	P99
TOS intranet domain name	113.555	102.514	137.896	163.218	251.795
TOS accelerated domain name	28.309	30.503	36.927	41.692	61.646

Last updated: 2026.08.25 10:41:51