memory leak of virtual memory

By -

I run tensorflow on linux (ubuntu20). TF executes my c++ functions for graph compilation/destruction.

The consumption of the process virtual memory grows until out-of-memory (>40GB) and the process is killed.

I track malloc/free and mmap/munmap with LD_PRELOAD hook and compare with the process virtual memory consumption from /proc/self/status (VmSize). Each graph compilation increases both malloc-allocated and the process virtual memory by almost the same size.

Graph destruction decreases malloc-allocated size but not the process virtual memory.

So in spite of malloc-allocated memory staying overall stable, the process virtual memory grows fast.

e.g.:

before compile: 41MB[mmap]/3320MB[malloc]/12428MB[process]
after  compile: 46MB[mmap]/7434MB[malloc]/16529MB[process]
before destroy: 46MB[mmap]/7436MB[malloc]/16593MB[process]
after  destroy: 46MB[mmap]/3250MB[malloc]/16593MB[process]

graphDestroy does not destroy everything by design so a small leftover is expected.

I tried to play with mallopt(M_MMAP_THRESHOLD) with no result.

What else can be done in order to find the leak?

UPDATE:

i add here steps that i tried and a way that worked - maybe it can be useful to someone.

the functions themselves are tested with sanitizers in unit tests. valgrind crashes the app before the main training loop starts. So this direction was a dead end.

i wanted to collect memory stats from glibc. unfortunately mallinfo is useless, mallinfo2 is not available on ubuntu20 and malloc_info prints too much. so i tried to use jemalloc and its function malloc_stats_print for stats.

the stats looked ok. but app behavior changed - virtual memory still grew (up to 75GB) but the resident memory stayed stable (~20GB) and the app worked with no memory issues.

then i tried to run my app without jemalloc but with a periodic call of malloc_trim(0) and it behaved the same way as jemalloc (virtual memory grew but resident memory stayed stable).

conclusion: sometimes malloc_trim can fix an issue that looks like a leak.

a good article about it



Comments

Post a Comment

Popular posts from this blog

Today Walkin 14th-Sept

By -
Image
14th Sept 2017 ( GO Through all List) Read Walkin details properly , Try to call company before attending if possible !!! ==================================== 1. Myntra through fresherworld (Call Based , Venue shared to shortlisted candidates) _____________________________ 2. Magna : Hiring for Client Harman : C++
Read more

Spring Elasticsearch Operations

By -
Elasticsearch Operations Spring Data Elasticsearch uses two interfaces to define the operations that can be called against an Elasticsearch index. These are ElasticsearchOperations and ReactiveElasticsearchOperations. Whereas the first is used with the classic synchronous implementations, the second one uses reactive infrastructure. The default implementations of the interfaces offer: Read/Write mapping support for domain types. A rich query and criteria api. Resource management and Exception translation. ElasticsearchTemplate The ElasticsearchTemplate is an implementation of the ElasticsearchOperations interface using the Transport Client. @Configuration public class TransportClientConfig extends ElasticsearchConfigurationSupport {   @Bean   public Client elasticsearchClient() throws UnknownHostException {                      Settings settings = Settings.builder().put("cluster.name", "elasticsearch")....
Read more

Hibernate Search - Elasticsearch with JSON manipulation

By -
Elasticsearch with JSON manipulation Elasticsearch ships with many features. It is possible that at some point, one feature you need will not be exposed by the Search DSL. To work around such limitations, Hibernate Search provides ways to: Transform the HTTP request sent to Elasticsearch for search queries. Read the raw JSON of the HTTP response received from Elasticsearch for search queries. Direct changes to the HTTP request may conflict with Hibernate Search features and be supported differently by different versions of Elasticsearch. Similarly, the content of the HTTP response may change depending on the version of Elasticsearch, depending on which Hibernate Search features are used, and even depending on how Hibernate Search features are implemented. Thus, features relying on direct access to HTTP requests or responses cannot be guaranteed to continue to work when upgrading Hibernate Search, even for micro upgrades (x.y.z to x.y.(z+1)). Use this at your own risk....
Read more