standards icon indicating copy to clipboard operation
standards copied to clipboard

Add larger memory flavors

Open scoopex opened this issue 7 months ago • 12 comments

The openstack flavor manager fetches the flavor definitions from the conformance test. These definitions contain a "mandatory" and a "recommended" section.

I think it would be a good idea to have a set of default flavors in this file with the larger amounts of memory described, because in my opinion there is a high probability that something like this will be used very often (e.g. for database page caches). On the other hand, I don't think that variations of larger numbers of CPUs (apart from the existing ones) are needed very, because huge cpu workloads are often better distributed by multiple virtual machines (i.e. by using kubernetes).

I therefore advocate that we make flavors with larger amounts of RAM available in the recommended section of that file, as otherwise there will be more variants or SCS users will have to contact their provider every time.

From my point of view, it is very useful for SCS users if they can also expect larger flavors from different suppliers.

scoopex avatar May 25 '25 16:05 scoopex

As discussed on Matrix (https://matrix.to/#/!xlFjWCAoQenuOeIVQw:matrix.org/$2bVMQjnP0QpGkOVtlVSusyt4gTQfaoGybYPkBUn2u2M?via=matrix.org&via=regio.chat&via=matrix.uhurutec.com) and suggested by @mbuechse: this is a good candidate to being discussed in the next Lean Coffee. The next Lean SCS Operator Coffee is on June 12.

fkr avatar May 26 '25 06:05 fkr

As discussed in the Lean SCS Operator Coffee on June 12, a survey of the SCS members and market analysis may be a good way to find out which additional (larger) default flavors are needed.

@garloff recommended to add the new flavors as recommended first and make them mandatory later to give providers time to adapt.

depressiveRobot avatar Jun 13 '25 14:06 depressiveRobot

Converted this to draft because this is mainly a point for organizing the discussion. If the larger flavors are to be added, then the file(s) to be changed would be a little different.

mbuechse avatar Jun 25 '25 11:06 mbuechse

Hello there.

I ran some statistics:

COUNT   RAM     VCPUS  RATIO
    144 4096    2      2:1
    109 1024    1      1:1
     64 8192    4      2:1
     52 65536   16     4:1
     40 2048    1      2:1
     39 16384   8      2:1
     38 32768   8      4:1
     31 16384   4      4:1
     28 4096    1      4:1
     27 32768   16     2:1
     22 8192    2      4:1
     16 512     1      0.5:1

This is our current use in the currently most relevant one of our public cloud sites. Grouped by ratio:

COUNT    RATIO
    314  2:1
    171  4:1
    109  1:1
     16  0.5:1

Not sure if I can make it to this week's standardisation SIG, but there we go.

All these flavors are dedicated thread (not dedicated core and not shared core).

Needless to say, this cloud is not SCS compliant, also because of the flavors. And obviously, this is biased, as our customers can only use what we offer. Hence, there cannot be any usage of >=8:1 flavors or of flavors with more than 64 GiB of RAM in this particular site.

(We are working on renewing the public cloud offering and are of course looking into the general direction of SCS flavors for inspiration of how we build it.)

horazont avatar Jun 30 '25 08:06 horazont

Further discussion in SIG Std/Cert meeting on 2025-07-10 with the following suggestion:

  • add larger memory flavors as "recommended" with 64 and 128 GiB of RAM for CPU/RAM ratios 1:2, 1:4, 1:8
  • brief explanation what "recommended" means

What would be the best place to add a brief explanation of what we mean by “recommended”?

depressiveRobot avatar Jul 11 '25 13:07 depressiveRobot

If you want to offer flavors only for specific memory sizes (e.g. 32 GB), you should include all recommended flavors for this specific memory size.

"Recommended" does not mean that you should use all available flavors from the entire list. Instead, the recommendation only refers to the flavors for your desired configuration (e.g. 32 GB).

Example: If you want to offer 32 GB instances, only include the flavors recommended for 32 GB - not all flavors from 2 GB to 128 GB that might be available in total.

berendt avatar Jul 11 '25 14:07 berendt

If you want to offer flavors only for specific memory sizes (e.g. 32 GB), you should include all recommended flavors for this specific memory size.

"Recommended" does not mean that you should use all available flavors from the entire list. Instead, the recommendation only refers to the flavors for your desired configuration (e.g. 32 GB).

Example: If you want to offer 32 GB instances, only include the flavors recommended for 32 GB - not all flavors from 2 GB to 128 GB that might be available in total.

Thank you for the short and precise explanation.

Another thing which popped into my head: Are SCS-32V-64, SCS-32V-128 and SCS-64V-128 even desired flavors?

depressiveRobot avatar Jul 11 '25 14:07 depressiveRobot

From my perspective this is good to go.

garloff avatar Oct 26 '25 17:10 garloff

Whats the story with this?

fkr avatar Nov 18 '25 13:11 fkr

It has to be verified that this change is consistent after the most recent commits by @garloff. It was a topic for the ALASCA Summit Hackathon, but we unfortunately didn't get around to it.

mbuechse avatar Nov 18 '25 13:11 mbuechse

I had a look. It seems fine.

mbuechse avatar Nov 20 '25 15:11 mbuechse

@fkr I think this should be reviewed by someone who's not me, garloff or scoopex.

mbuechse avatar Nov 20 '25 15:11 mbuechse

@fkr I think this should be reviewed by someone who's not me, garloff or scoopex.

@garloff

mbuechse avatar Nov 24 '25 15:11 mbuechse

@fkr I think this should be reviewed by someone who's not me, garloff or scoopex.

I concur - github added you automatically because of the CODEOWNERS file... :)

fkr avatar Nov 24 '25 18:11 fkr

@garloff @berendt @chrisschwa please comment and approve (or not)

fkr avatar Nov 24 '25 18:11 fkr

From my perspective, this is good to go. But as main contributor, I'm not the best person to judge ...

garloff avatar Dec 01 '25 10:12 garloff

LGMT

chrisschwa avatar Dec 01 '25 11:12 chrisschwa