1.20.0-beta.1 regression in memory usage

[carols10cents]

This is going to be vague; please feel free to close if this is covered by more specific issues (I wasn't sure) or if this is too vague to be useful.

I've seen intermittent failures with Rust 1.20.0-beta.1 (and sometimes nightly as well, but never stable) on travis when testing crates.io. The failures appear to be because it has run out of memory; this job in particular failed with:

= note: /usr/bin/ld: final link failed: Cannot allocate memory

Other samples:

• https://travis-ci.org/rust-lang/crates.io/jobs/262872876
• https://travis-ci.org/rust-lang/crates.io/jobs/262872877

It seems to happen more often on this PR. The particular job linked above is crates.io master plus the tar crate added as a dependency.

Changing the travis config to sudo: required (which apparently gives you more memory) gets that PR to pass on beta, so I'm not particularly blocked, but just wanted to raise this in case it's a useful canary in the coal mine.

[arielb1]

= note: "cc" "-Wl,--as-needed" "-Wl,-z,noexecstack" "-m64" "-L" "/home/travis/.rustup/toolchains/beta-x86_64-unknown-linux-gnu/lib/rustlib/x86_64-unknown-linux-gnu/lib" "/home/travis/build/rust-lang/crates.io/target/debug/deps/server-9f3e8bea00ce5d30.0.o" "-o" "/home/travis/build/rust-lang/crates.io/target/debug/deps/server-9f3e8bea00ce5d30"

This looks like linking the final executable - I don't think there should be any Rust code running in parallel.

[alexcrichton]

• -Z time-passes for the stable and beta compilers show nothing major that changed for the server binary.
• The target directory grows by ~100MB from stable to beta, with the whole thing being on the order of a gigabyte.o.
• Link time for beta and stable is about the same hovering around 4-5s locally using ld.bfd.
• massif shows the peak memory usage of the linker is ~270MB for stable and ~290MB for beta

Nothing looks too out of place? Maybe the travis machines were just getting unlucky?

[sfackler]

It seems a bit strange that there's 10% more stuff being generated by beta compared to stable, right?

[carols10cents]

As discussed in the core team meeting, @alexcrichton was testing on crates.io master, whereas I was seeing pretty consistent failures on rust-lang/crates.io#869 (without the workaround commit I added to that PR that switches travis from sudo: false to sudo: required. Alex said he would redo his tests on that PR ❤️

[alexcrichton]

I tested for that PR as well but unfortunatel all the data is the same. There's very little difference between beta/stable on crates.io w/ that codebase. My only real guess here is that crates.io was hovering around some memory limit previously and just happened to blow it this time around

[carols10cents]

My only real guess here is that crates.io was hovering around some memory limit previously and just happened to blow it this time around

@wycats seemed to indicate we should have been seeing more intermittent failures over time if that was the case; these cropped up suddenly.

That PR was adding a few more crates?

Do you think we should just close this? I'd love to track memory usage of crates.io compilations in travis over time, but it feels nontrivial to set up....

[alexcrichton]

Since crates.io builds on stable it's not totally crazy that we might be able to add it to perf.rust-lang.org! It may take some time to fully integrate it and get good timings though. In the meantime though I think I'm at least personally out of ideas, although we could always leave the regression open to see if others would like to give it a stab

[Marwes]

Seeing this in gluon as well. https://travis-ci.org/gluon-lang/gluon/jobs/266597837. Is enabling sudo the only workaround for now?

[alexcrichton]

@Marwes the error there is "No space left on device", is that related to this issue?

[Marwes]

@alexcrichton Hmm, I guess that would be disk space actually. Started seeing various crashes when linking in beta and figured that this issue would be related.