Advanced search

Message boards : Number crunching : performance loss through patches against Meltdown and Spectre

Author Message
Erich56
Send message
Joined: 1 Jan 15
Posts: 1120
Credit: 8,967,195,176
RAC: 31,162,806
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwat
Message 48650 - Posted: 6 Jan 2018 | 17:49:14 UTC

what I am very curious about is how much the CPU performance loss through the patches against Meltdown and Spectre, which for Windows will be published shortly, and for Linux allegedly have been distributed already, will affect GPUGRID crunching.

Any guess by the experts here?

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 1332
Credit: 7,092,067,459
RAC: 15,079,789
Level
Tyr
Scientific publications
watwatwatwatwat
Message 48652 - Posted: 6 Jan 2018 | 22:05:24 UTC - in response to Message 48650.

So far, the tests I have seen show almost no performance loss in typical computer workloads. Some performance loss, about 3% in synthetic benchmarks for certain tests. Most of the performance loss will be in the datacenters.

I have no idea what performance loss there might be for applications here as I have no idea what kind of code they run. I doubt there will be any since I believe the compute loads are purely mathematical and there is little I/O for any task. It is in I/O that they are seeing some degradation after the patch.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48664 - Posted: 7 Jan 2018 | 14:07:57 UTC - in response to Message 48650.
Last modified: 7 Jan 2018 | 14:16:07 UTC

patches against Meltdown and Spectre
FYI: there are at least one kind of Spectre attacks, which in theory (therefore in practice) can't be fixed by software patches, as these attacks exploit the effects on the cache of the "out of order execution" scheme of the CPU (which is hardwired to the chip). It means that these CPUs are basically doomed. There are mayor breakdowns ahead of us in every area of our lives.

Erich56
Send message
Joined: 1 Jan 15
Posts: 1120
Credit: 8,967,195,176
RAC: 31,162,806
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwat
Message 48665 - Posted: 7 Jan 2018 | 14:56:00 UTC - in response to Message 48664.

...It means that these CPUs are basically doomed. There are mayor breakdowns ahead of us in every area of our lives.

nice to read :-)

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48667 - Posted: 7 Jan 2018 | 16:26:30 UTC - in response to Message 48665.

...It means that these CPUs are basically doomed. There are mayor breakdowns ahead of us in every area of our lives.
nice to read :-)
From the smiley I think that the severity of this bug is simply too unthinkable, but one app reading the other app's memory without any software or hardware could detect it is simply the end of any kind of cyber security. No passwords, financial records, PIN codes, nothing is safe from now on. That is basically the end of the world as we know it.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48669 - Posted: 7 Jan 2018 | 17:30:06 UTC - in response to Message 48667.

That is basically the end of the world as we know it.

Does this mean that "out of order execution" can't be used at all, or just needs to be implemented differently?

It sounds like a good opportunity for AMD and Intel to sell more chips.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48670 - Posted: 7 Jan 2018 | 17:57:00 UTC - in response to Message 48669.

Does this mean that "out of order execution" can't be used at all, or just needs to be implemented differently?
There are CPUs without "out of order execution" which are safe - those are the slowest, cheapest ones (some older Atom CPUs for example), so the first part is true. The present implementation of the "out of order execution" with a little addition (=new CPUs) would be safe from the present implementation of Spectre attacks, but the present architecture thought to be safe for more than 20 years. No one can tell how long it takes to "hack" the modified architecture as well.

It sounds like a good opportunity for AMD and Intel to sell more chips.
Provided that there will be anybody to buy it.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48671 - Posted: 7 Jan 2018 | 18:38:44 UTC - in response to Message 48670.

Provided that there will be anybody to buy it.

It is a strange way for civilization to end: The Out of Order Apocalypse.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1617
Credit: 8,111,694,351
RAC: 18,129,731
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48673 - Posted: 7 Jan 2018 | 22:13:49 UTC

Reposting from SETI:

Spectre vulnerability affects Nvidia cards

Make sure you update your drivers.

https://nvidia.custhelp.com/app/answers/detail/a_id/4611/related/1

It isn't just CPUs, apparently.

Erich56
Send message
Joined: 1 Jan 15
Posts: 1120
Credit: 8,967,195,176
RAC: 31,162,806
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwat
Message 48679 - Posted: 9 Jan 2018 | 14:13:19 UTC

I am just reading something rather worrrying in the LHC forums, with regards to the Meltdown/Spectre patches:

"Details of a problem have been gradually emerging, People with AMD Athlon-powered computers say that following the installation of the patch, it is impossible to boot into Windows leaving a full re-installation as the only option -- although some users report that even this does not fix the problem."
(possibly related to the antivirus program incompatibility)

https://betanews.com/2018/01/08/microsoft-meltdown-spectre-patch-bricks-amd-pcs

3de64piB5uZAS6SUNt1GFDU9d...
Avatar
Send message
Joined: 20 Apr 15
Posts: 285
Credit: 1,102,216,607
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwat
Message 48681 - Posted: 9 Jan 2018 | 18:43:35 UTC - in response to Message 48671.

It is a strange way for civilization to end: The Out of Order Apocalypse.


Well, it is alarming indeed. On the other hand, we should not overrate Spectre. If one (malignant) application can access some memory content of another by reading the pre-fetch, it basically means that you must have that malware on your computer already. And if so, you could already do a lot of other bad (but already existing) things, which are cheaper to implement, installing keyloggers for example. So Spectre is not necessarily the end of days ... but it's the large attack surface which makes it grievous and will cause us some headache for the next couple of months.

Well, I guess both NSA and KGB know a few gaps other than that, which are not known yet but also inherent to the architecture. And if a powerful organisation really wants to know your online banking password, they will contrive ways and means, be sure about that! All that protects us ordinary people is our unimportance and relatively low bank balance. It is a cold comfort, I know.

____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48683 - Posted: 9 Jan 2018 | 21:34:51 UTC - in response to Message 48681.

If one (malignant) application can access some memory content of another by reading the pre-fetch, it basically means that you must have that malware on your computer already. And if so, you could already do a lot of other bad (but already existing) things, which are cheaper to implement, installing keyloggers for example.

Sure. The AV software should offer us ordinary users the usual degree of protection. Whether it is so easy to protect servers I don't know, but they will think of something.

Erich56
Send message
Joined: 1 Jan 15
Posts: 1120
Credit: 8,967,195,176
RAC: 31,162,806
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwat
Message 48684 - Posted: 11 Jan 2018 | 5:55:01 UTC

Some further information:

On the GPU front we can see that since cache pre-fetch is the issue that all classes of GPU/CPU & other processor class with cache may well face issues.

https://insidehpc.com/2018/01/nvidia-races-patch-gpu-drivers-spectre-meltdown/

https://nvidia.custhelp.com/app/answers/detail/a_id/4611

Profile JStateson
Avatar
Send message
Joined: 31 Oct 08
Posts: 186
Credit: 3,361,527,550
RAC: 283,434
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 48805 - Posted: 3 Feb 2018 | 17:11:52 UTC

The only problem I have had related to those possible threats is that the ubuntu "fix" for it in kernel 4.13.0.32 broke the nVidia CUDA OpenCL package. I had to revert back to my original 4.13.0.19.generic to get my CUDA cards crunching again. I found a complaint at the nVidia github where developers complained that too many changes went into the 32 kernel which caused their code to fail.

Post to thread

Message boards : Number crunching : performance loss through patches against Meltdown and Spectre

//