Building tensorflow serving with TCMalloc

[jeongukjae]

Feature Request

Describe the problem the feature is intended to solve

Recently I found that building tensorflow serving with tcmalloc and set soft limit can mitigating these kind of memory issue: https://github.com/tensorflow/serving/issues/2142 , https://github.com/tensorflow/serving/issues/1664. Also, I could get slightly better performances.

Here's what I did.

• build tensorflow_model_server cc binary with tcmalloc using malloc argument.
• launch background thread before server starts using tcmalloc::MallocExtension::ProcessBackgroundActions (reference)
• Set soft limit, and add tcmalloc_soft_limit argument for it

Describe the solution

How about providing tensorflow serving compiled with tcmalloc? I know I can use jemalloc instead, but jemalloc has so much configurations, while tcmalloc doesn't. It is easy to use and has great performance.

[singhniraj08]

@jeongukjae,

Thank you for filing this feature request. We will discuss this feature implementation internally and update this thread. Thanks.