[yocto] Suitable machine for yocto

[Mark Hatle]

On 9/10/17 2:31 PM, Alex Lennon wrote:
> 
> 
> On 10/09/2017 19:17, Mark Hatle wrote:
>> On 9/10/17 11:14 AM, Alex Lennon wrote:
>>>
>>> On 10/09/2017 17:06, Mark Hatle wrote:
>>>> On 9/10/17 2:00 AM, Usman Haider wrote:
>>>>> Hi,
>>>>>
>>>>> Can someone please recommend some good machine for yocto environment and
>>>>> building sdks. I am interested in RAM, hard disk space, processor.
>>>> You want fast I/O, as much RAM and as many (fast) cores as you can afford.  I
>>>> don't think there is a single answer as what is 'best'.  It also depends on
>>>> which Yocto Project versions, and which layers you are using as to which
>>>> combination is best.
>>>>
>>>> I run builds on my laptop, 4-core/8-thread & SSD and 16 GB of ram from a few
>>>> years ago.  It's fast, but I wouldn't want to do all of my development on it.
>>>>
>>>> I've had 8-core/16-thread (32GB ram/standard disk), 16-core/32-thread (72GB
>>>> ram/SAS-3 RAID), 24-core/48-thread (64GB ram/SATA - software RAID), 72-core/144
>>>> thread (256 GB ram/hardware raid/SAS-3), and recently upgraded to
>>>> 96-core/192-thread (256 GB ram/hardware raid/SAS-3).
>>>>
>>>> I would not go below quad-core (8-thread) myself.  You can get a quad core, good
>>>> quality machine for $1000 or less these day.  If you move up to the larger
>>>> machines, you can even be able to get to a 24-core for less then $5000.  By the
>>>> time you get to 96-core and all of the googles you are likely talking $50000 or
>>>> more.
>>>>
>>>> By clock raid, the 24-core machine is the fastest..  While the 96-core monster
>>>> can do the builds the quickest.  But when you figure out cost/performance/etc..
>>>> the 24-core is probably the best performance per dollar, and with adequate RAM
>>>> (I'd say at least 64GB if not 128GB), and fast I/O you'll probably get the
>>>> lowest price for the best performance in that category.
>>>>
>>>> If you need sheer speed and price is no option, then the (4 CPU w/ 24 core each)
>>>> 96-core monster (or even better) is what you want to go with.  256GB ram would
>>>> be a minimum with that configuration (I'm not sure if more is actually helpful,
>>>> I rarely end up in swap -- but I go get into situations where more then 50% of
>>>> ram is used.)  With that many cores, disk I/O starts to become obvious.  So
>>>> faster the better... SSDs would be the fastest, but of course the most expensive.
>>>>
>>>> If your employer is paying for the machine, you may be able to get a better then
>>>> normal machine by explaining how much time a faster machine will save and how
>>>> comparing to your salary a machine is inexpensive.  (If you are a contractor or
>>>> student, that changes of course.)  :)
>>>>
>>>> So my point is really, figure out how much money you have to spend.  My rule of
>>>> thumb is roughly:
>>>>
>>>> 1) Buy as many cores as you can.  Try to get a CPU that has Hyperthreading or
>>>> equivalent to double the effective core count.  Fastest processing speed helps
>>>> in repetitive cases vs full system builds.
>>>>
>>>> So if the choice is a 24 core @ 2.2GHz vs 22 core @ 2.5, I'd probably go with
>>>> the 22-core.  While if it was 24 core @ 2.2GHz vs 8 core @ 4.2 GHz, I'd go with
>>>> the 24 core.
>>>>
>>>> 2) Try to get at least 1 GB of ram per thread (2 GB per core..)  You can cut
>>>> back on the ram (if necessary) once you hit 72 threads or so.   (72 threads
>>>> right now seems to cover most of the parallelization in a full system build.
>>>> There are points in the system where it can parallelize MUCH more, but they are
>>>> fairly rare.)
>>>>
>>>> 3) You need fast disks.  Software RAID works fine, but you likely need to buy at
>>>> least a couple of disk to boost performance.  SSDs are fast, but lots of builds
>>>> take space, so fast SATA or even better SAS drives are the best performance per
>>>> cost.
>>> This brings to mind a related question I keep coming back to as to the
>>> economics of a docker (or similar) image running a fast Yocto build in
>>> the cloud.
>>>
>>> i.e. set config params -> bring up server image on plaform A/B/C ->
>>> perform build taking time X/Y/Z -> store output images -> bring down
>>> server  == $ ?
>>>
>>> I find myself asking what the optimal cost per-build would be using this
>>> approach...
>> I helped someone do some very -preliminary- figured a few years ago.  The
>> processing was 'cheap', but between storage and network transfer costs.. it was
>> cheaper to buy a reasonable machine.. payback time was only a few months.
>>
>> (cloud 'storage' is often very slow as well, because there are expectations of
>> migration and things like that.)
>>
>> So as of a few years ago at least, the economics didn't factory the cloud -- yet.
>>
>>
> 
> Yeah that is more or less my thinking. Although I do imagine that if a 
> source mirror on an internal network leg could be set up then that might 
> significantly reduce chargeable network b/w...

premirror for downloads and sstate-cache shared between multiple customers could
reduce the network b/w far enough to make it economical.   But you would need to
trust other vendors since the sstate-cache can be manipulated to provide
malicious software (in some cases).  (And strengthening the sstate-cache
checking can cause enough extra overhead to make it 'cheaper' to build then use
sstate-cache in many cases.)

--Mark

> Cheers,
> 
> Alex
>