Power

Hello,

since POWER10 do not support active memory sharing, what is the best practice to dynamic optimize memory between the LPARs?

Best regards,

P. Tanovski

This is a very good question - I have no quick fix but a few suggestions:

• Moving memory between LPARs is is a great trick but you also need to consider moving memory between servers. No I do not mean yanking out DIMMs etc. But in a cluster you may have one server that is bottlenecked on RAM and one bottlenecked on CPU.  LPM could be used to free up RAM on one server for LPARs can grow their foot print with dynamic memory allocation changes.
• It is relatively easy to determine LPARs with a lack of memory - just watch the paging. Although, with large LPARs and experience you will know large LPAR can satisfactorily page all day due to processes changing the memory residence sets during the day. Assuming, the paging space is across many devices.
• Detecting LPARs with lots of spare memory is easy - just look for Free Memory in the configuration.
• Detecting LPARs that have used all memory once via the buffer pool for IO may have memory that is not actively in use - it is just not on a free list. Showing a little of my AIX bias here.  In this case there is the "old rmss trick" which can reveal this non-Free memory that is really spare.
• All this requires monitoring and I have my nmon and njmon tools in mind.  But the reviewing of the stats is still person-power intensive.
• Perhaps, this might make a good AI project.. If the stats are gathered by njmon + InfluxDB (timeseries database) for easy real time access the memory review could be centralised.
• Servers these days have far far larger memory sizes than the "good old days" when I started in IT. Don't get me talking about the computer I used with only 128 bytes of RAM - and I don't mean KB.  It this an "old timer" problem - i.e. trying to optimise expensive resources where a set and forget approach covers 99% of the problem.
• Perhaps, it is not so important - just purchase more RAM is a way to fix performance issues due to a lack of memory.  The CMS tools allow system admin to switch on and off memory across a cluster of servers.  So you can license-wise move memory towards your CPU intensive but low RAM size LPARs.