Fusion and RT Rreempt [Was: RTAI/fusion 0.7.2]

[Philippe Gerum]

Max Krasnyansky wrote:
> Philippe Gerum wrote:
> 
>>> Here is something even more interesting.
>>> I just ran the same test on the IBM Z-Pro machine:
> 
>  >> [snip]
> 
>>> latency -p 4000 + high network load
>>> RTS|        1262|      780892|     1724248|       0|    
>>> 00:00:19/00:00:19
>>>
>>> How about that :) ? 1.7 milliseconds. Wow.
>>> This is totally reproduceable btw. I played with the BIOS settings 
>>> and stuff, makes no difference. MSI workaround is enabled but makes 
>>> no difference.
>>>
>>
>> If you look at the average figure, it says that the pathological 
>> latency is almost continuously happening (> 780 us), so the problem is 
>> likely related to this specific hw there. It would be interesting to 
>> know whether running a kernel with IO-APIC and ACPI disabled exhibits 
>> the same problem.
> 
> Compiling :)
> 
>>> I was wondering what kind of numbers are you guys seeing on IA64 
>>> machines ?
>>>
>>
>> 4-way (1.3Ghz) Bull Novascale servers currently give ~50 us worst-case 
>> for an earlier version of fusion used by the HYADES project that 
>> contributed this port. UP gives ~30 us.
> 

> I see. That still kinda suxx :)

Well, yes and now. SMP adds a fair amount of intrinsic latency, and 
those figures have been obtained with a test equivalent to 
testsuite/latency, i.e. user-space. This said, you would not want an 
ia64-based server in your living room anyway, it's definitely too noisy...

> I need an architecture that gives ulta 
> low latencies (say under 10 usec) and loads of CPU horse power. Looks 
> like dedicated
> CPU (keep the cache hot, etc) is the only option.

Likely, yes. Even with a very smart latency compensation mechanism, I 
don't see how x86 or ia64-based archs could fit such stringent requirement.

-- 

Philippe.