Add more POWER8 optimizations

[Mauricio Faria de Oliveira]

There are some upstream commits that introduce optimizations with significant performance improvement for IBM POWER8.
This patch introduces their minimal bits/changes in the source package.

[Mauricio Faria de Oliveira (mfo)]

Here is the comparison / benchmarking of:
1) current package
2) current package + attached debdiff
3) upstream code

It shows that:
1) performance of current package improves a *lot* with the attached debdiff;
2) performance of current package + attached debdiff is equal to the upstream's.

$ openssl speed -evp aes-128-gcm
--------------------------------

1) current package

 $ openssl speed -evp aes-128-gcm
 <...>
 The 'numbers' are in 1000s of bytes per second processed.
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-gcm 80062.35k 87264.06k 89732.78k 90357.90k 90450.60k

2) current package + attached debdiff

 $ openssl speed -evp aes-128-gcm
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-gcm 244362.49k 558183.91k 1240341.76k 1534709.55k 1625676.37k

 speedup:
 - 16 bytes: 3x
 - 64 bytes: 6.4x
 - 256 bytes: 13.8x
 - 1024 bytes: 17x
 - 8192 bytes: 18x

3) upstream code

 $ LD_LIBRARY_PATH=/usr/local/ssl/lib/ /usr/local/ssl/bin/openssl speed -evp aes-128-gcm
 <...>
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-gcm 244700.26k 562547.05k 1243287.37k 1535474.69k 1620669.78k

 (same as #2)

$ openssl speed -evp aes-128-ctr
--------------------------------

1) current package

 $ openssl speed -evp aes-128-ctr
 <...>
 The 'numbers' are in 1000s of bytes per second processed.
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-ctr 144150.81k 154729.39k 157233.48k 158093.92k 157728.99k

2) current package + attached debdiff

 $ openssl speed -evp aes-128-ctr
 <...>
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-ctr 345134.48k 842190.28k 2709975.21k 4183632.82k 4959908.05k

 speedup:
 - 16 bytes: 2.4x
 - 64 bytes: 5.4x
 - 256 bytes: 17.2x
 - 1024 bytes: 26.4x
 - 8192 bytes: 31.4x

3) upstream code

 $ LD_LIBRARY_PATH=/usr/local/ssl/lib/ /usr/local/ssl/bin/openssl speed -evp aes-128-ctr
 <...>
 type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
 aes-128-ctr 345758.55k 839220.54k 2694016.68k 4168019.39k 4974437.20k

 (same as #2)

[Mauricio Faria de Oliveira (mfo)]

In summary:

$ openssl speed -evp aes-128-gcm
--------------------------------

 speedup:
 - 16 bytes: 3x
 - 64 bytes: 6.4x
 - 256 bytes: 13.8x
 - 1024 bytes: 17x
 - 8192 bytes: 18x

$ openssl speed -evp aes-128-ctr
--------------------------------

 speedup:
 - 16 bytes: 2.4x
 - 64 bytes: 5.4x
 - 256 bytes: 17.2x
 - 1024 bytes: 26.4x
 - 8192 bytes: 31.4x

[Mauricio Faria de Oliveira (mfo)]

In the patches, this line might not be as clear as intended:

> The 'backport' keyword in 'Origin:' just indicates it isn't equal to 'upstream'.

"isn't equal" in the sense only the interesting changed files were kept, but not changed, and other files were removed from the diff.

[Ubuntu Foundations Team Bug Bot (crichton)]

The attachment "openssl-power8_part2.debdiff" seems to be a debdiff. The ubuntu-sponsors team has been subscribed to the bug report so that they can review and hopefully sponsor the debdiff. If the attachment isn't a patch, please remove the "patch" flag from the attachment, remove the "patch" tag, and if you are member of the ~ubuntu-sponsors, unsubscribe the team.

[This is an automated message performed by a Launchpad user owned by ~brian-murray, for any issue please contact him.]

[Marc Deslauriers (mdeslaur)]

Subscribing the release team to see if this is too late to get a feature exception.

[Adam Conrad (adconrad)]

I'd be inclined to say it's too late in the cycle to appropriately validate this on all architectures, especially given the partial backports that very much touch common code, etc.

I think it's saner to defer this to 15.10, either with more complete backports or "for free" from a new upstream release, and get solid testing in on all 6 arches.

[Mauricio Faria de Oliveira (mfo)]

Hi Adam,

> I'd be inclined to say it's too late in the cycle [...] given the partial backports that very much touch common code, etc.

That's certainly understandable.

I believe it's possible to trim up the backports even more.

AFAICT, most changes are contained inside #ifdef HWAES_CAPABLE blocks, and the first patch is just syntactic sugar so not to change a few lines in subsequent patches; so I'll try to remove it, and change those few lines (the 'ks' to 'ks->ks' change).

Then it should be OK to validate the places defining HWAES in other files/arches..
or even restrict any addition/ifdef even more w/ a POWER8/ppc64el specific conditional.

Does that help?

Thanks

[Mauricio Faria de Oliveira (mfo)]

Ok, the trimmed backports:
- removed the patch which touched common code w/out #ifdef blocks.
- now all changes are under #ifdef blocks, which helps analysis.

Build log:

 ...
 ALL TESTS SUCCESSFUL.
 make[2]: Leaving directory '/home/ubuntu/ss5/openssl-1.0.1f/test'
 ...

In the patches, all changes are under these macros; now, checking for their
definitions will identify if any other arches might be affected.

 1)
 #ifdef HWAES_CAPABLE
  #ifdef HWAES_cbc_encrypt
  #ifdef HWAES_ctr32_encrypt_blocks

 2)
 #if ... && (... || defined(OPENSSL_CPUID_OBJ))

 3)
 #ifdef HWAES_CAPABLE
  #ifdef HWAES_ctr32_encrypt_blocks

[Mauricio Faria de Oliveira (mfo)]

Adam,

Does it help/improve the situation at all?

[Mauricio Faria de Oliveira (mfo)]

Well, a later SRU should be OK too, right? :)

[Sebastien Bacher (seb128)]

Thanks, it looks like those changes are in the wily version of openssl right? Closing the bug but feel free to reopen if I'm wrong

[Mauricio Faria de Oliveira (mfo)]

Thanks, Sebastien. Yes, the changes are in wily.
For documentation purposes. The numbers in a KVM guest on POWER8 match those in comment #2.

$ lsb_release -d
Description: Ubuntu Wily Werewolf (development branch)

$ dpkg -s openssl | grep ^Version:
Version: 1.0.2d-0ubuntu1

$ openssl speed -evp aes-128-gcm
...
aes-128-gcm 245286.05k 562263.25k 1246015.74k 1527520.94k 1611202.56k

$ openssl speed -evp aes-128-ctr
...
aes-128-ctr 343161.12k 846347.46k 2721303.72k 4169064.79k 4955196.07k