[Beowulf] [upgrade strategy] Intel CPU design bug & security flaw - kernel fix imposes performance penalty

[John Hearns]

I guess we have all seen this:   https://access.redhat.com/articles/3307751

If not, 'HPC Workloads' (*) such as HPL are 2-5% affected.
However as someone who recently installed a lot of NVMe drives for a fast
filesystem, the 8-19% performance hit on random IO to NVMe drives is not
pleasing.


(*) Quotes are deliberate.  We all know that the best benchmarks are your
own applications.

On 6 January 2018 at 12:05, John Hearns <hearnsj at googlemail.com> wrote:

> Disabling branch prediction  - that in itself will have an effect on
> performance.
>
> One thing I read about the hardware is that the table which holds the
> branch predictions is shared between processes running on the same CPU core.
> That is part of the attack process - the malicious process has knowledge
> of what the 'sharing' process will branch to.
>
> I float the following idea - perhaps this reinforces good practice for
> running HPC codes. Meaning cpusets and process pinning,
> which we already do for reasons of performance and for better resource
> allocation.
> I expose my ignorance here, and wonder if we will see more containerised
> workloads, which are strictly contained within their own memory space.
> I then answer myself by saying I am talking nonsense, because the kernel
> routines need to be run somewhere and this exploit is all about being able
> to probe
> areas of memory which you should not be able to do by speculatively
> running some instructions and capturing what effect they have.
> And ""their own memory space" is within virtual memory.
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
> On 6 January 2018 at 02:26, Christopher Samuel <chris at csamuel.org> wrote:
>
>> On 06/01/18 12:00, Gerald Henriksen wrote:
>>
>> For anyone interested this is AMD's response:
>>>
>>> https://www.amd.com/en/corporate/speculative-execution
>>>
>>
>> Cool, so variant 1 is likely the one that SuSE has firmware for to
>> disable branch prediction on Epyc.
>>
>> cheers,
>> Chris
>> --
>>  Chris Samuel  :  http://www.csamuel.org/  :  Melbourne, VIC
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit
>> http://www.beowulf.org/mailman/listinfo/beowulf
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20180107/1c241af2/attachment.html>