[nycbug-talk] Hifn card

[Isaac Levy]

On Jun 6, 2012, at 2:20 PM, George Neville-Neil wrote:
> On Jun 6, 2012, at 14:06 , Isaac Levy wrote:
>> Hi George,
>> 
>> On Jun 6, 2012, at 12:58 PM, George Rosamond wrote:
>>> On 06/06/12 12:48, George Rosamond wrote:
>>>> Wondering about others' experiences with Hifn card on FreeBSD.
>>>> 
>>>> Looking for something not too expensive, mostly for SSL/TLS acceleration.
>>>> 
>>>> Have used padlock(4) and glxsb(4), the later on Alix boards.
>>>> 
>>>> Went through the hifn(4) man page, but would love to hear about
>>>> experiences, not read about them.
>>> 
>>> Well still curious to hear about others' experiences, but it seems for what I need, CPU matters more, in that there will be lots of little work, not some big fat work :)
>>> 
>>> g
>> 
>> 2 things worth less than .02¢:
>> 
>> 1) Depends on your use, CPU has blown away what the little hifn cards can push.
>> (I have a 4 year old story about pulling them out of Soekris 5501's and getting measured 3-5x throughput increase for IPSEC VPN's- no kidding.  [was PFSense 1.2, which was of course, FreeBSD 7.x based])  Interrupts and data throughput to the card were killing network IO, and the CPU totally smoked the accelerator on it's own.  Stats fuzzy, never looked back after pulling the cards.
>> 
> 
> At the moment this is what I'd expect.  Unless you're on something with low power, like an old
> Soekris, then your CPU is probably fine.
> 
> More interesting is if people start doing crypto on GPU, but I'm not going to even think about that.


Ha!  Or CPU architecture could simply take those cues and start changing again altogether…  (I'm over-simplifying things, but isn't it about time?)

"Yeah. RISC is good."

Rocket-
.ike