Experiment: test: Disable fsync in travis tests

[laanwj]

Install the package "eatmydata" which disables fsync and friends, and use this while running the functional tests.

Not sure whether it applies to travis, but at least locally on a VM this made some of the tests run three times as fast.

[maflcko]

functional tests are disabled for windows and no_wallet, so I don't think we need this here.

You probably did this for consistency/simpler code.

[laanwj]

My intent was to add it for all the lines that have RUN_TESTS=true

[maflcko]

Concept ACK:

*** BEFORE ***
ALL                            | ✓ Passed  | 2648 s (accumulated) 
Runtime: 713 s

*** AFTER ***
ALL                            | ✓ Passed  | 1290 s (accumulated) 
Runtime: 337 s

As we ignore the return code of fsync in our code, this should not add additional debug hassle in case the fsync fails to write everything to disk (maybe due to hardware issues)?

Edit: The comparison is wrong. I accidentally compared a CRON job against this pull request, where this pull runs a subset of the tests of a cron job.

[laanwj]

As we ignore the return code of fsync in our code, this should not add additional debug hassle in case the fsync fails to write everything to disk (maybe due to hardware issues)?

I don't think so.

It does remind me that we could use the same technique (library interposing) to inject failures for testing to test how bitcoind handles i/o failures.

[laanwj]

Argh, it fails on 32-bit linux:

ERROR: ld.so: object '/usr/lib/libeatmydata/libeatmydata.so' from LD_PRELOAD cannot be preloaded (wrong ELF class: ELFCLASS64): ignored.

[laanwj]

I can't get it to work on 32-bit: not with 32-bit and not with 64-bit eatmydata.

The problem, I think, is that python runs as 64-bit, and bitcoind as 32-bit. So with the 64-bit eatmydata it fails on bitcoind, with the 32-bit one it fails on python.

Not sure what would be a good way to fix this. I suppose a hack that sets the environment variable BITCOIND so that only bitcoind is started through eatmydata (and not the outer script) could work.

Edit: BITCOIND="eatmydata $PWD/src/bitcoind" doesn't work, at least. This would need an actual wrapper script.

[laanwj]

Hm another idea, instead of using the wrapper/interposer, which seems to be tricky on mixed architctures, couldn't we just link bitcoind against libeatmydata.so directly?
On Ubuntu 16.04 libeatmydata1 is a separate package that can be installed, this isn't the case on trusty/14.04.

16.04 (just for illustration):

$ dpkg -L libeatmydata1
...
/usr/lib/x86_64-linux-gnu/libeatmydata.so.1.1.2
/usr/lib/x86_64-linux-gnu/libeatmydata.so.1
/usr/lib/x86_64-linux-gnu/libeatmydata.so

14.04 (which is what travis uses):

$ dpkg -L eatmydata
...
/usr/lib/libeatmydata/libeatmydata.so

[laanwj]

With the wrapper most works, but it still fails in p2p-versionbits-warning.py, as this makes bitcoind invoke another process which apparently is 64 bit:

p2p-versionbits-warning.py failed, Duration: 8 s
stdout:
2017-04-17 17:09:13.351000 TestFramework (INFO): Initializing test directory /tmp/testmahvm5qh/275
2017-04-17 17:09:13.609000 TestFramework.mininode (INFO): Connecting to Bitcoin Node: 127.0.0.1:13200
2017-04-17 17:09:19.543000 TestFramework (INFO): Stopping nodes
2017-04-17 17:09:21.665000 TestFramework (INFO): Cleaning up
2017-04-17 17:09:21.666000 TestFramework (INFO): Tests successful
stderr:
ERROR: ld.so: object '/usr/lib/libeatmydata/libeatmydata.so' from LD_PRELOAD cannot be preloaded (wrong ELF class: ELFCLASS32): ignored.
ERROR: ld.so: object '/usr/lib/libeatmydata/libeatmydata.so' from LD_PRELOAD cannot be preloaded (wrong ELF class: ELFCLASS32): ignored.

Sigh...

[jnewbery]

this makes bitcoind invoke another process

The only process I can think of that p2p-versionbits-warning.py specifically would invoke is echo:

        self.extra_args = [["-alertnotify=echo %s >> \"" + self.alert_filename + "\""]]

This causes bitcoind to launch a thread to run echo when it detects blocks signalling unknown versionbits.

[laanwj]

Hm another idea, instead of using the wrapper/interposer, which seems to be tricky on mixed architctures, couldn't we just link bitcoind against libeatmydata.so directly?

I've tried this locally with some success, but not entirely sure how to port it to travis. Trying...

[laanwj]

Oh wow that worked.

But seems @MarcoFalke 's comparison was mistaken and this doesn't actually make a difference so closing...

[maflcko]

@laanwj Sorry for my confirmation bias. On a first glance those numbers seemed perfectly plausible.

[laanwj]

Yes it would have been plausible; apparently Travis isn't using qemu/kvm or has already configured their VMs to not pass through fsync.