search

CCI-MOC / ops-issues

2 stars 0 forks source link

Pricing for different classes of bare metal machines shared under ESI #1271

Closed msdisme closed 3 months ago

msdisme commented 5 months ago

Motivation

We require pricing for bare metal machines so that we can start to include bare metal in our offerings.

Completion Criteria

A pricing guide for bare metal decided upon.

Description

Completion dates

Desired - 2024-04-17 Required - TBD

tzumainn commented 5 months ago

Related issue: https://app.zenhub.com/workspaces/moc-a--rh-sprint-workspace-62a210f69d42f600151deae0/issues/gh/cci-moc/esi/509

naved001 commented 5 months ago

Related issue on how we want to classify machines in ESI. We should be on the same page https://github.com/CCI-MOC/esi/issues/509

joachimweyl commented 5 months ago

@msdisme at the base level this is just (purchase price + expenses)/5/360/24 but I assume we have more complexity we plan to include?

msdisme commented 4 months ago

The propose a cost for fc430 and fc830 should be sooner so we may communicate it to RH OKDMOCA so they may consider including in FY2025 budget.

joachimweyl commented 4 months ago

@joachimweyl provide a price based on SU count * CPU SU Cost for FC430 & FC830

joachimweyl commented 4 months ago
RAM (GB) Threads Ψ SU Price/hr Unused
FC430 96 24 24 $ 0.312 16 threads
FC430 128 32 32 $ 0.416 8 threads
FC430 * 40 40 limited by RAM $ 0.52 SU not possible
FC830 2,048 144 144 $ 1.872 1,472 GB
FC830 2,048 288 288 $ 3.744 896 GB
FC830 2048 * 512 $6.656 SU not possible

FC430 - 128 GB RAM / 96 GB RAM Threads 20*2=40, 24 SU for 96 GB RAM and 32 SU for 128 GB RAM with some unused CPU 16 for 96 GB RAM and 8 for 128 GB RAM or if you calculate by threads it is 40 SU 24 SU = $0.312/hr 32 SU = $0.416/hr 40 SU = $0.52/hr

FC830 - 2,048 GB RAM Threads 36*4= 144 SU, with a lot of unused RAM 1,472 GB or with high ram SU 288 SU, with some unused RAM 896 GB, or if you calculate by RAM 2,048/4 = 512 SU 144 SU = $1.872/hr 288 SU = $3.744/hr 512 SU = $6.656/hr

Clearly, we can't charge full SU without either RAM or Threads to back it up but we need to charge for the extra resource so it is somewhere between the values provided. maybe in the high $3 range for the FC830 and .39 for the 96 RAM 430 and .45 for the 128 RAM?

msdisme commented 4 months ago

FC430: 96 at .35 FC430: 128 at .45 FC830: evaluate vs. commercial compare to commercial pricing for all of these, as well (discount against that.)
price for equivalent machine to price them out

msdisme commented 4 months ago

re. commercial pricing look at aws, google, and a few of the lower cost providers of bare metal: Liquid Web Cherry Servers Hivelocity Scaleway Amazon INAP RedSwitches OpenMetal Oracle Vultr

joachimweyl commented 4 months ago

FC830 comprable BM Machines

Company Name RAM (GB) Threads Price $/hr Price if Enough threads Price if enough ram Average
Amazon x2idn.metal 2048 128 13.338 15.00525 13.338 14.171625
OpenMetal none 1024 128 2.784 3.132 5.568 4.35
Vultr none 512 64 2.344 5.274 9.376 7.325
Vultr none 2048 256 8.185 4.6040625 8.185 6.39453125
Google m3-megamem-128 1952 128 17.0033 19.1287125 17.83952787 18.48412018
Google n2-highmem-128 864 128 10.080725 11.34081563 23.89505185 17.61793374
Azure Ms-Type1 2048 128 16.576 18.648 16.576 17.612
Cherry Servers AMD EPYC 9554P 1152 128 2.456 2.763 4.366222222 3.564611111

Still looking up

  1. Liquid Web - highest threads is 64 no real comparison
  2. Hivelocity - max memory is 256GB
  3. Scaleway - max threads is 64
  4. INAP - max threads 48
  5. RedSwitches - max threads 64

@msdisme would you prefer this in google sheets so you can sort it?

msdisme commented 4 months ago

thanks, turns out cut and paste from github goes smoothly into google sheets so we are good.

joachimweyl commented 4 months ago

Based on this data it looks like around $3.789 would be a fair price since none of the low-end BM listed in the table have as much RAM and the higher-end options are clearly higher than we want to go. Also keep in mind what we could make off of this node if we sold it as a bunch of high-ram SUs we would be making closer to $6.656/hr. @msdisme thoughts?

joachimweyl commented 4 months ago

@joachimweyl add FC430 commercial comparison as a new table.

joachimweyl commented 4 months ago

FC430 comparable BM machines

Company Name RAM (GB) Threads Price $/hr Price if Enough threads Price if enough ram for 96 Price if enough ram for 128 Average for 96 Average for 128
Vultr none 128 16 0.521 1.3025 0.39075 4.168 0.846625 2.73525
Vultr none 256 48 1.079 0.8991666667 0.404625 2.877333333 0.6518958333 1.88825
Google n2-standard-32 128 32 2.032486 2.5406075 1.5243645 8.129944 2.032486 5.33527575
Google n2-standard-16 64 16 1.016243 2.5406075 1.5243645 8.129944 2.032486 5.33527575
Cherry Servers none 64 32 0.8986 1.12325 1.3479 3.5944 1.235575 2.358825
Cherry Servers none 256 48 1.0486 0.8738333333 0.393225 2.796266667 0.6335291667 1.83505
joachimweyl commented 4 months ago

Another way to look at it is in google sheets. Based on the tables I would suggest FC430 (96GB) - $0.35-$0.65 FC430 (128GB) - $0.45-$0.75 FC830 - $2.67 - $3.97

msdisme commented 3 months ago

seem reasonable - I will communicate this to the folks using it even if we are not yet ready to bill for it.

joachimweyl commented 2 months ago

@msdisme just to confirm bare metal pricing for GPU systems are calculated by multiplying the number of GPUs in the node by the SU price for that GPU, correct?