Object Storage Service:Performance metrics of OSS accelerator

Operations on large amounts of data

In the following example, you use ossutil to download or read 10,000 objects whose size is 100 KB (total object size: 976 MB) by using an OSS internal endpoint and the accelerated endpoint of an OSS accelerator.

• Test plan

  Test tool	Operation	Description
  ossutil	Run the cp command to download 10,000 objects whose size is 100 KB from an OSS bucket to your local computer.	Use the OSS internal endpoint and the accelerated endpoint of the accelerator to test the speed at which objects in the OSS directory are concurrently downloaded to the local computer.

• Test results

  Test tool	Use the OSS internal endpoint	Use the OSS accelerator for data preloading and the accelerated endpoint of the OSS accelerator for accelerated access
  ossutil	2.2 MB/s	24 MB/s

• Test conclusion

  • Downloading data by using ossutil when the OSS accelerator feature is enabled is approximately 10 times faster than downloading data by using ossutil when the OSS accelerator feature is disabled.

  The preceding example describes how the OSS accelerator feature can significantly improve the data transmission and access speed when you use tools, such as ossutil, to perform batch operations on large amounts of data.

Machine learning or deep learning for large amounts of data

In the following example, you read data from OssIterableDataset and OssMapDataset datasets created by OSS Connector for AI/ML by using an OSS internal endpoint and the accelerated endpoint of an accelerator. 10,000,000 objects whose average size is 100 KB (total object size: 1 TB) are used.

• Test parameters

  Parameter	Value/Operation	Description
  dataloader batch size	256	Each batch task processes 256 samples.
  dataloader workers	32	Data is loaded in parallel by using 32 processes.
  transform	def transform(object): data = object.read() return object.key, object.label	Data is not preprocessed.

• Test results

  Dataset created by using	Dataset type	Use the internal endpoint	Use the OSS accelerator for data preloading and the accelerated endpoint for accelerated access
  OSS Connector for AI/ML	OssIterableDataset	99920 img/s	123043 img/s
  	OssMapDataset	56564 img/s	78264 img/s

• Test conclusion

  Reading data from OssIterableDataset and OssMapDataset datasets with the OSS accelerator feature enabled is approximately 1.6 times faster than reading the same data with the OSS accelerator feature disabled. OSS Connector for AI/ML can handle highly concurrent access at a high bandwidth level when the OSS accelerator is disabled. Using OSS Connector for AI/ML together with the OSS accelerator feature provides even more powerful performance.

Download response latency statistics

Download objects that are 10 MB in size multiple times for testing and calculate the response latency in milliseconds when you disable and enable OSS accelerator. When you disable OSS accelerator, objects are downloaded from OSS.

In the following figure, P50 indicates that 50% of requests meet the current latency statistics, and P999 indicates that 99.9% of requests meet the current latency statistics.

The results show that the latency is reduced by 10 times when an OSS accelerator is used.

Data lakes and data warehouses in the cloud

A user tests the performance when a local disk, OSS, and an OSS accelerator are used as storage media. Approximately 2 billion pieces of data in a lineitem table whose size is 760 GB is used.

• Latency

  Scenario	Local CacheFS (local disk)	OSS	OSS accelerator
  Point queries	382ms	2451ms	1160ms
  Random queries on 1,000 data items	438ms	3786ms	1536ms
  Random queries on 10% of data	130564ms	345707ms	134659ms
  Full scan	171548ms	398681ms	197134ms

• Performance

  • During online queries, the response time of the OSS accelerator is 2 to 2.5 times higher than the response time of OSS. During full scan and random queries on 10% of data, the performance of the OSS accelerator is 2 to 2.5 times higher than the performance of OSS and 85% of the performance of local ESSD CacheFS.

  • During online queries, the fixed latency of a single request sent to the OSS accelerator is 8 to 10 ms. During random queries on 1,000 data items and point queries, the performance of the OSS accelerator is 1.5 to 3 times higher than the performance of OSS and 30% of the performance of local ESSD CacheFS.

Simulation training for containers and autonomous driving

When you enable the OSS accelerator feature, a large number of containers are started at the same time to obtain images, maps, and log data and the overall duration of simulation training is reduced by 60%.