Effort-less way to run C++ clients

[qiranq99]

Hi,

I found it quite troublesome to get a C++ client all the way through compiling the entire project and trying to understand the raw C++ APIs in the doc, thus wondering is there a easier way to access the C++ clients, e.g., via partial compilation.

BTW, have you ever profiled the performance gap between a Python client and a C++ client regarding the client-server IPC overhead (inside one machine) when retrieving data via data = client.get(oid)? According to our experiments, each retrieval takes several hundreds of microseconds (using AMD EPYC 7763), which is unacceptable in some latency-sensitive workloads, which leads to 2 questions:

1. would a C++ client outperforms a Python one in terms of IPC overhead?
2. would v6d be optimized for low latency scenarios in the future?

[sighingnow]

Hi @qiranq99,

Thanks for touching!

I found it quite troublesome to get a C++ client all the way through compiling the entire project and trying to understand the raw C++ APIs in the doc, thus wondering is there a easier way to access the C++ clients, e.g., via partial compilation.

Yes, we do have a set of CMake options to control which components to enable: https://github.com/v6d-io/v6d/blob/main/CMakeLists.txt#L56-L65

If you only need the C++ client to access metadata and blobs, you could just enable the client and disable all other components.

BTW, have you ever profiled the performance gap between a Python client and a C++ client regarding the client-server IPC overhead (inside one machine) when retrieving data via data = client.get(oid)?

As it uses shared memory and avoids any potential data copies. The cost should be a very small constant and won't scale with the size of your data.

ccording to our experiments, each retrieval takes several hundreds of microseconds (using AMD EPYC 7763)

May I know more about the test case (maybe some code snippets that I can use to reproduce the performance gap). We would investigate to check if there are any regression.

would a C++ client outperforms a Python one in terms of IPC overhead?

There should be no performance differences about if Python is used or not. Let us know if you have encountered such problems.

would v6d be optimized for low latency scenarios in the future?

v6d is mainly optimized for sharing big data objects (e.g., tensors, tables, dataframes) between processes.

[qiranq99]

Hi @sighingnow,

import vineyard
import numpy as np

client = vineyard.connect()

data = np.arange(100000)
oid = client.put(data)

# %%timeit
retrieved_data = client.get(oid)

basically we took the above example as a benchmark, and the tested latency is several hundreds of microseconds, while some low-latency-oriented object store could deliver several tens of nanoseconds. Though v6d is not aiming for low latency, <50us latency of retrieving data from the server is what we expect.

[qiranq99]

BTW, I successfully compiled C++ client in isolation thanks to your hint.

However, the C++ API reference seems to be generated directly from mkdoc tools and there is no clear guidance on how to use the C++ library.

[sighingnow]

basically we took the above example as a benchmark, and the tested latency is several hundreds of microseconds, while some low-latency-oriented object store could deliver several tens of nanoseconds.

Will investigate. In my queue now.

However, the C++ API reference seems to be generated directly from mkdoc tools and there is no clear guidance on how to use the C++ library.

More user-friendly tutorials about the C++ APIs are in our roadmap. For now, you may refer to our unittests as examples for usage: https://github.com/v6d-io/v6d/tree/main/test

Sorry for the inconvenience.

[qiranq99]

Benchmark update:

on getting a bytearray(10) from local object stores,

• v6d: ~500 us
• Ray Plasma: ~100 us
• Redis: ~20 us (Redis is fast only for small objects)

Machine: Intel Xeon 4316 @ 2.30GHz

[github-actions[bot]]

/cc @sighingnow, this issus/pr has had no activity for for a long time, could you folks help to review the status ? To suppress further notifications,

• for issues,
  • if it is waiting for further response from the reporter/author, please help to add the label requires-further-info,
  • if you have already started working on it, please add the label work-in-progress to the issue,
  • if this issue requires further designing discussion and not in current plan, or won't be fixed, please add the label requires-further-discussion or wontfix to the issue,
• for pull requests,
  • if you are still working on it and it is not ready for reviewing, please convert this pull request as draft PR,
  • if you have decided to hold this development on, please add the requires-further-discussion label to the pull request. Thanks!

[github-actions[bot]]

/cc @sighingnow, this issus/pr has had no activity for for a long time, could you folks help to review the status ? To suppress further notifications,

• for issues,
  • if it is waiting for further response from the reporter/author, please help to add the label requires-further-info,
  • if you have already started working on it, please add the label work-in-progress to the issue,
  • if this issue requires further designing discussion and not in current plan, or won't be fixed, please add the label requires-further-discussion or wontfix to the issue,
• for pull requests,
  • if you are still working on it and it is not ready for reviewing, please convert this pull request as draft PR,
  • if you have decided to hold this development on, please add the requires-further-discussion label to the pull request. Thanks!