[Info-vax] Data Center Efficiency (was: Re: Disable hyper-threading from the DCL prompt)

[Stephen Hoffman]

On 2015-03-06 21:28:50 +0000, JF Mezei said:

> Question:
> 
> Will the move to x86 require VMS add support for fancy CPU features 
> that can be turned on/off like hyperthreading ?

Baseline power management and hyperthreading support are already 
available on OpenVMS I64.   This as JF already knows.

> Operating systens liek OS-X seem to have plenty of automated features 
> for energy saving, core managemnt, shutting down cores when not needed 
> etc.  Wondering if VMS will HAVE to support any of those, or "SHOULD" 
> surpport any of those or whether those are more marketing PR gimmicks 
> that are not actually needed in real data centres ?

Duh.  But then JF already knows this is a "duh", of course.

Data centers are often interested in lowering their power and cooling 
requirements, which means that mechanisms for reducing power and for 
powering down some or most of a server can be of interest.  OS X and 
iOS do have integrated, automated mechanisms for saving power, though 
these are for general efficiency and battery life — in practical terms, 
longer battery life isn't usually a central goal that most data centers 
are interested in, but sipping power and lowering cooling loads 
definitely is.  
<http://en.wikipedia.org/wiki/Green_computing#Data_center_power> The OS 
X and iOS mechanisms use various approaches, such as seeking to 
coalesce interrupts across multiple applications to allow the cores to 
stay powered down longer, fast suspend and restarts, as well as other 
techniques that seek to keep the system and the processors and the 
peripherals in lower-powered states or even powered down.

What is coalescing interrupts?  If you have a dozen active processes 
and each with, for instance, scheduled interrupts at one-second 
intervals, then having all of those interrupts hit in quick succession 
across one or more cores can mean that the processor cores can power 
down for longer, as differentiated from having those same dozen 
one-second interrupts scattered throughout each second across 
potentially random cores, and leaving the entire processor and its 
various cores and the memory caches that can be present or can be 
flushed in each core in a state analogous to flailing.

Certainly for cases requiring absolute server and application 
responsiveness, programmers and system administrators don't want to see 
peripherals in low-powered states that need to be quickly accessed, but 
for environments where the associated latency of powering up devices is 
less central, then the power savings can be interesting.  Some folks 
might not want to see some or all of their disks spun down for 
instance, but others will want the power and cooling savings that can 
be realized from powering down parts of a rotating-rust array, and 
other folks might well be well along migrating to solid-state storage 
and which is both more efficient and usually powers up much faster than 
rotating rust.

<http://www.energystar.gov/ia/partners/prod_development/downloads/DataCenters_GreenGrid02042010.pdf>, 
starting around page 14.
<http://www.energystar.gov/ia/partners/prod_development/downloads/EPA_Report_Exec_Summary_Final.pdf> 

<https://www.energystar.gov/ia/products/power_mgt/downloads/Kaiser_Permanente_case_study.pdf> 

<http://www.energystar.gov/ia/products/downloads/ES_server_case_study.pdf>
<http://www.energystar.gov/ia/products/power_mgt/ES_Data_Center_Utility_Guide.pdf> 

etc.

The application and the server operating system software is a core part 
of this discussion, as it is the software that increasingly controls 
the energy-reducing features of many of the current-generation 
computers, and some or all of the future-generation computers.

Wrapping this back to OpenVMS specifically, the Itanium and Alpha boxes 
are — for the amount of local use that they can get at many sites — 
much less efficient than x86-64 servers, due to the technologies used 
in the older designs, and due to the sheer size and expandability of 
the boxes involved.  A two-core i2- or i4-class Itanium is a huge box.  
Some folks need that, but some don't and can seek to consolidate 
(not-OpenVMS) environments using VMs or Docker containers or similar 
mechanisms.   Consolidation is a given, and going to smaller boxes, and 
the boxes are only getting smaller and/or denser[1].   The associated 
costs of power and cooling are only going to increase, too.   Put 
another way, it's a completely different world for OpenVMS now, and 
power and cooling are always going to be a factor.

______
[1] A Mac Pro would be a decent-sized Alpha, in terms of the numbers of 
cores (up to 12 cores) and the maximum memory (up to 64 GB) supported, 
and I'd tend to expect the Mac Pro to be somewhere between faster and 
much faster than most Alpha boxes, and faster than at least the 
early-generation Itanium boxes, too


-- 
Pure Personal Opinion | HoffmanLabs LLC