Re: ldom question

Sorry, hit reply initially, not reply all.


On Fri, Apr 10, 2015 at 9:57 AM, Scarlett <[email protected]> wrote:

> On 10/04/2015 08:36, Henning Brauer wrote:
>
>> * Scarlett <[email protected]> [2015-04-07 19:32]:
>>
>>> On 07/04/2015 15:12, Jeremiah Ford wrote:
>>>
>>>> My last question, a perhaps less constructive one, why isn't this more
>>>> widely used?  There are so many people running OpenBSD on Linux KVM, I
>>>> would think that this is the solution they have been looking for.
>>>> Thanks again, already read the manual for ldomctl, plan on finishing
>>>> reading Sun's docs next.
>>>>
>>> Lack of mindshare is the primary reason.
>>>
>>
>> Well. The sparc64 ldoms are dog slow compared to the usual amd64
>> virtualization options, with a comparable power budget and hardware
>> cost rather much higher than lower. I wish it was different :(
>>
>> The static config requiring a host/dom0/$preferred_nomenclature reboot
>> to reassign CPUs etc is also hindering - but the above really is the
>> killer.
>>
>>  For most cheap VPS providers Linux is considered the primary target, with
>>> the exception of a few like RootBSD, where FreeBSD is the focus. It'd be
>>> nice if there was someone providing OpenBSD-backed SPARC64, though.
>>>
>>
>> We use ldoms for some services internally, writing the systems
>> integration for automatic config/deployment, start/stop, console access
>> etc never made sense from the business PoV.
>>
>> <shameless plug>
>> if you really want a sparc64 ldom, poke me and I'm sure we can work
>> sth out.
>> </shameless plug>
>>
>>
> That's a shame :/ Do you know/have any guesses where the performance
> problems are coming from?
>
>
The T1000/2000 computers had a single socket CPU that was a 4, 6 or 8
cores, 4 thread/core, clock speed of 1.0Ghz, 1.2Ghz or 1.4Ghz (this tended
to be an 8 core box with 64GB or 128GB of RAM, back in the day)   A thread
was about the speed of a 400Mhz CPU, so it was really good for web servers
and oracle databases, but it's horrible for straight-line performance.  It
also only had FP unit per socket, so if you're were doing a lot of FP, it
was really bad.

The T-2 chipset (T5440 and that range) was a 1.6Ghz clock, 8 threads/core,
1 FP per core (not per socket as above), and you could get up to 4 sockets
in a box. A thread is about a 600-700Mhz CPU.

I think the T-4 chipset is a 2.8Ghz clock, 8 threads/core, 1 FP per core,
up to 4 sockets per box.  This CPU has a dynamic CPU scheuderl.  We have a
pair and are running IO Domain LDOMs for Databases.  For one of our
databases, with the primary running the LDOM 2.0 (Oracle OVM), we have the
threading model set to max-ipc, which allows us to collasce the cores so
they appear a 8 straight line cpus, instead of 64 vCPUs.  The other
database uses the normal database "max-throughput" threading model.  A
thread is about a 700-900Mhz CPU.  In max-ipc, we see 2.8Ghz cores reported.

The T-5 chipset is a 3.4 or 3.6Ghz cores, 16 threads/core, dynamic CPU
model.  In LDOM 3.0, Oracle took away the ability to change the threading
model so I don't think a  T5 can be forced to the max-ipc model anymore.

Hope that explains why a 10 year old computer doesn't have very good
performance. ;-)

>
> Mostly I just want something SPARC64 and small of my own, to use for
> testing software and maybe working on a few ports. For the second use a
> local box with radeondrm support would be more convenient, though...
>

Regards,

Ben