java very slow on arm

[Orion Poplawski]

I'm building protobuf 3.1.0 for rawhide.  The arm build is incredibly slow.

See http://koji.fedoraproject.org/koji/taskinfo?taskID=16512668

Entire x86_64 build took 21 minutes. Steps in the isValidUtf8Test look like:

Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 4:18:02 PM com.google.protobuf.IsValidUtf8TestUtil testBytes

It's been running for over 11.5 hours on arm, and test steps look like:

Nov 18, 2016 6:06:31 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 6:07:50 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 6:09:15 PM com.google.protobuf.IsValidUtf8TestUtil testBytes

and has increased in time to:

Nov 19, 2016 2:52:51 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Processed 15 million characters
Nov 19, 2016 3:15:25 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Processed 16 million characters
Nov 19, 2016 3:32:52 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 2650112 of 16777216

Have there been any other reports of issues like this on arm?

[Bill Chatfield]

My understanding is that it is slow because there is no JIT for arm. Please correct me if I'm wrong. I ran Java on Debian on my Mac Powerbook (PPC) when the JIT wasn't working and it was unusable.

On Friday, November 18, 2016 10:49 PM, Orion Poplawski  wrote:

I'm building protobuf 3.1.0 for rawhide.  The arm build is incredibly slow.

See http://koji.fedoraproject.org/koji/taskinfo?taskID=16512668

Entire x86_64 build took 21 minutes. Steps in the isValidUtf8Test look like:

Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 4:18:02 PM com.google.protobuf.IsValidUtf8TestUtil testBytes

It's been running for over 11.5 hours on arm, and test steps look like:

Nov 18, 2016 6:06:31 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 6:07:50 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 18304 of 65536
Nov 18, 2016 6:09:15 PM com.google.protobuf.IsValidUtf8TestUtil testBytes

and has increased in time to:

Nov 19, 2016 2:52:51 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Processed 15 million characters
Nov 19, 2016 3:15:25 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Processed 16 million characters
Nov 19, 2016 3:32:52 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
INFO: Round tripped 2650112 of 16777216

Have there been any other reports of issues like this on arm?

[Orion Poplawski]

On 11/19/2016 07:44 PM, Bill Chatfield wrote:
> My understanding is that it is slow because there is no JIT for arm. Please correct me if I'm wrong. I ran Java on Debian on my Mac Powerbook (PPC) when the JIT wasn't working and it was unusable.

Yeah, that's what I ultimately decided as well.  It does look like there is
some progress upstream on this, so hopefully it will speed up at some point.
> On Friday, November 18, 2016 10:49 PM, Orion Poplawski  wrote:
>
> I'm building protobuf 3.1.0 for rawhide.  The arm build is incredibly slow.
>
> See http://koji.fedoraproject.org/koji/taskinfo?taskID=16512668
>
> Entire x86_64 build took 21 minutes. Steps in the isValidUtf8Test look like:
>
> Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Round tripped 18304 of 65536
> Nov 18, 2016 4:18:01 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Round tripped 18304 of 65536
> Nov 18, 2016 4:18:02 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
>
> It's been running for over 11.5 hours on arm, and test steps look like:
>
> Nov 18, 2016 6:06:31 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Round tripped 18304 of 65536
> Nov 18, 2016 6:07:50 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Round tripped 18304 of 65536
> Nov 18, 2016 6:09:15 PM com.google.protobuf.IsValidUtf8TestUtil testBytes
>
> and has increased in time to:
>
> Nov 19, 2016 2:52:51 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Processed 15 million characters
> Nov 19, 2016 3:15:25 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Processed 16 million characters
> Nov 19, 2016 3:32:52 AM com.google.protobuf.IsValidUtf8TestUtil testBytes
> INFO: Round tripped 2650112 of 16777216
>
> Have there been any other reports of issues like this on arm?

[Severin Gehwolf]

On Fri, 2016-11-18 at 20:49 -0700, Orion Poplawski wrote:
> I'm building protobuf 3.1.0 for rawhide.  The arm build is incredibly slow.

Yes, the 32 bit arm build of java-1.8.0-openjdk is using a Zero JVM. No
JIT. So yes, it's slow. As an alternative you can try the java-1.8.0-
openjdk-aarch32 package[1]. I'm not sure how stable that is, though.

Cheers,
Severin

[1] http://koji.fedoraproject.org/koji/buildinfo?buildID=812478

[Jiri Vanek]

On 11/21/2016 10:20 AM, Severin Gehwolf wrote:
> On Fri, 2016-11-18 at 20:49 -0700, Orion Poplawski wrote:
> > I'm building protobuf 3.1.0 for rawhide.  The arm build is incredibly slow.
>
> Yes, the 32 bit arm build of java-1.8.0-openjdk is using a Zero JVM. No
> JIT. So yes, it's slow. As an alternative you can try the java-1.8.0-
> openjdk-aarch32 package[1]. I'm not sure how stable that is, though.

It is stable enough. And its main purpose is koji builder.
In rawhide we are already providing java-devel in this package. So generally build root can get one 
from java-1.8.0-openjdk java-1.8.0-openjdk-aarch32 packages in random.

Maybe we can remove java-devel provides from java-1.8.0-openjdk on arm32 in rawhide ?
J.