[AusNOG] Anyone know of VPN's being bandwidth managed (throttled)

[Mark Smith]

On 8 January 2017 at 12:50, Peter Tiggerdine <ptiggerdine at gmail.com> wrote:
> I would think most people are using CPU software encryption these
> days.

I showed that CPU software encryption these days is likely to be a lot
fast than you might think. My 2009 era CPU is capable of 105 MiB/s or
around 800 Mbps crypto throughput.

If you're specifically going to use crypto today, the price of
hardware crypto is one of these cheap Intel compute sticks for around
$120 to $150.

http://ark.intel.com/products/91065/Intel-Compute-Stick-STK1AW32SC?_ga=1.134383985.1003419038.1470461640

http://ark.intel.com/products/87383/Intel-Atom-x5-Z8300-Processor-2M-Cache-up-to-1_84-GHz

> How many firewalls come with crypto cards out of the box?
>

If your firewall is using an Intel CPU from the past 5 years, there is
very likely an AES engine on the CPU.

People need to realise that special things, if there is enough demand
for it, become normal, out-of-the-box things, as it is one of the ways
a vendor can make their product more attractive over a competitor's.

If people aren't enabling crypto because they assume the overhead is
too high, without making any current observations, measurements or
calculations, then their viewpoint is stuck in the 1990s and 2000s.

Crypto was special 15 years ago, however Intel and AMD have made it a
standard feature of most general purpose CPUs, including ones in
laptops, and low end ones such as Atoms and Celerons. Hardware crypto
is no where near as special and expensive as it used to be. For many
devices, if it isn't available to use, it's a software availability or
licensing problem.

https://ark.intel.com/search/advanced?AESTech=true


> However 1/3 of the bandwidth sounds like lots of RST or drops to the
> flow as tcp tries scale (which in turn makes the UDP packet bigger).
>
> I have to admit this is where LRO and LSO make it harder to troubleshoot.
>
>
> Regards,
>
> Peter Tiggerdine
>
> GPG Fingerprint: 2A3F EA19 F6C2 93C1 411D 5AB2 D5A8 E8A8 0E74 6127
>
>
> On Sun, Jan 8, 2017 at 11:36 AM, Mark Smith <markzzzsmith at gmail.com> wrote:
>> On 6 January 2017 at 13:38, Peter Tiggerdine <ptiggerdine at gmail.com> wrote:
>>> My experience is is that PMTUD isn't configured correctly (or not
>>> allow to pass along the path).. Also endpoint CPU. encrypting and
>>> decrypting is heavy on CPU (assuming you have no offload engine).
>>
>>
>> You might be surprised how much encryption throughput you get these
>> days when encrypting/decrypting in software on a CPU.
>>
>> On platforms that run openssl, you can measure crypto throughput for
>> various algorithms using 'openssl speed'.
>>
>> I've found that CPU software crypto under Linux on a 2009 era Intel
>> Q6600 is fast enough that I can encrypt the filesystems on that
>> machine and not notice the performance impact. Measuring using
>> 'cryptsetup benchmark' shows worst case throughput of 105.1 MiB/s for
>> aes-xts, which is what Fedora uses for filesystem crypto. My HDDs (WD
>> 2TB Blacks) do 123.58 MB/sec according to 'hdparm -t', so I am
>> sacrificing some HDD performance but it isn't noticeable.
>>
>> I wonder how many people are aware that modern Intel and AMD CPUs have
>> hardware AES crypto engines in them (known as 'AES NI' instructions)?
>> My 2013 Dell XPS laptop with an Intel Core i5 in it, using 'cryptsetup
>> benchmark', does worst case 1040 MiB/s aes-xts encryption/1060.1 MiB/s
>> decryption, which is more than twice the 426.53MB/s the SSD can do
>> (and around 10 times what my Q6600 does in software), so filesystem
>> crypto is not going to be the IO bottleneck.
>>
>> Using simple 8 bits per byte maths, that means my 2013 laptop could do
>> more than 8 Gbps of CPU crypto engine throughput. So if the CPU's AES
>> engine is used for VPN crypto, most people aren't going to have any
>> crypto throughput problems with a VPN (Peter Löthberg's grandma being
>> a possible exception).
>>
>> (I encrypt all my filesystems these days, not because I'm paranoid and
>> think the government is out to get me, but because it is easier to
>> know that if any one of my computers is stolen, they've stolen my
>> hardware but not my data.)
>>
>> Regards,
>> Mark.