1. Task 36b Scope – CPU & Memory (L=ChrisH)
Purpose – add CPU and Memory functionality support to the model
Specifically –
Includes – DSP, FPGA, GPU ? L1 & L2 cache, Page Files ?
Excludes – CPU and Memory physical inventory, which is covered by the existing Equipment model. Excludes Storage
External Dependencies –
Assumptions –
Risks –

2. Team Members
Leader - Chris Hartley
Members
???

3. IPR Declaration Is there any IPR associated with this presentation NO
NOTICE: This contribution has been prepared to assist the ONF. This document is offered to the ONF as a basis for discussion and is not a binding proposal on Cisco or any other company. The requirements are subject to change in form and numerical value after more study. Cisco specifically reserves the right to add to, amend, or withdraw statements contained herein.
THE INFORMATION HEREIN IS PROVIDED “AS IS,” WITHOUT ANY WARRANTIES OR REPRESENTATIONS, EXPRESS, IMPLIED OR STATUTORY, INCLUDING WITHOUT LIMITATION, WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

4. Need for a CPU and Memory Model
This is part of the data center triad :
Compute
Network  Done
Storage  Started
Information
Transfer
Modification N Y
Storage
Network
Compute
---

5. Why do we care about all this ?
The key questions that need to be supported are
Is my application working ?
Can users connect to my application ?
Managing each separately is easier, but doesn’t help if the issue is a mismatch between the compute and network worlds
Applications now are quite different than they were in 1981
Applications now have quite different network requirements than they had in 1981
1981 is an arbitrary date – the year the IBM PC was released

6. We need to support various arrangements of data and instructions
Here Processing Unit = CPU core (or FPGA, DSP, GPU … equivalent)
Single / Multiple
Instruction streams
Data streams

7. We need to support various arrangements of CPU and memory

8. Types of Memory have different speed and capacity tradeoffs
Increasing Capacity
(Not to scale)
CPU
Register
Cost per bit also is a factor
L1 Cache
Decreasing Speed + Latency
L2 Cache
Non-volatile
Main Memory
Storage
Volatile

9. DMTF CIM 2.48

10. DMTF CIM 2.48

11. SNIA Swordfish / DMTF Redfish
The key Redfish concepts appear to be :
Processor - a processor attached to a System
Memory - a Memory and its configuration
MemoryChunks - a Memory Chunk and its configuration
MemoryDomain - used to indicate to the client which Memory (DIMMs) can be grouped together in Memory Chunks to form interleave sets or otherwise grouped together

12. Openconfig

13. Proposed Model – Package Dependencies
Compute
Equipment
Core

14. Proposed CPU Model
If a CPU is symmetric then use one physical entry per CPU.
If a CPU is asymmetric, then use one physical entry per group of similar CPU cores
For example a CPU may have cores ( 4 * 1.8 GHz Type-A GHz Type-B) and hence 2 entries , each of 4 cores would be used
CPU chip(s) may be FRU or non FRU inventory – here we are interested in the compute function

15. Proposed Memory Model
Memory chip(s) / SIMM DIMM modules may be FRU or non FRU inventory – here we are interested in the memory function

16. (copied from Storage pack)
Compute Pool
Note that the decision was made to have a single compute pool rather than separate Storage, CPU and memory pools, because
CPU and memory are usually tightly coupled and the pool can then allocate these consistently
Sometimes storage is tightly coupled with CPU and memory and the pool can then allocate these consistently
Storage, CPU and memory may not be the best grouping of the pool entries – perhaps physical vs logical, input vs output – starting with a flat set of all the pool entries will allow the best grouping to evolve, rather than imposing an inheritance structure at this stage
Note that the pools aren’t hierarchical (deliberately no ComputePool contained in self-join)
The associations XxxPoolEntryIsLogical allow an output from one pool to become the input of another pool
We really want this to form a directed acyclic graph (no loops)

17. Some Questions
What units should we use for memory sizes – bytes, kiB, MiB … ?
What units should we use for CPU - Hz, kHz, MHz … ?
Note that Kubernetes works in units of CPU, where “One CPU, in Kubernetes, is equivalent to a Hyperthread on a bare-metal Intel processor with Hyperthreading”
A CPU hardware thread is also called a vCPU
Note that we should try and avoid using float numbers as the rounding proves problematic

18. Compute Pool Group
This will allow us to group for instance CPU and Memory that need to be kept together
Need to validate this

19. The Compute Model relates to the software using it

20. Compute Examples – Simple Host (1)
1 CPU with 1 core, 1 Memory DMM,
1 local HDD
CD = Phy(Chassis)
CD = Phy(Blade1)
CD = Phy(Blade2) PC
RunningOS
Running Software Process
OS invokes Process
PC from RunningSoftware Process
CPU Cores
Memory Blocks

21. Kubernetes Cluster Service …
Kubernetes
“A Pod represents a unit of deployment: a single instance of an application in Kubernetes, which might consist of either a single container or a small number of containers that are tightly coupled and that share resources” (eg. main + sidecar).
“Each Pod is meant to run a single instance of a given application.”
A Kubernetes Pod can request CPU and Memory and also have CPU and Memory Limits set
Memory is specified in bytes
“The CPU resource is measured in CPU units. One CPU, in Kubernetes, is equivalent to:
1 AWS vCPU
1 GCP Core
1 Azure vCore
1 Hyperthread on a bare-metal Intel processor with Hyperthreading
Fractional CPU values are allowed.” Precision finer than 1 milliCPU is not allowed.
Cluster
Service
Note that CPU usage is really policed as the usage (cycles – (milli)seconds of use in a period.
Assume we have a limit of 0.1 CPU and we check every 100ms.
Assume our process is using 50% CPU.
After 100ms we see it has used 50ms of CPU.
It is allowed only 10ms of CPU per 100ms (0.1 CPU) , so we pause it for 400ms – so it has used 50ms CPU in 500ms which is 0.1 !
Then we can allow it to run for another 100ms
OF course the CPU will likely vary and the scheduling algorithm will adapt for that.
Use Namespaces to split a cluster into ‘virtual’ Clusters
Container …

22. https://en.wikipedia.org/wiki/Symmetric_multiprocessing
Compute Examples – SMP

23. Compute Examples – CPU split into Cores

24. Compute Examples – Memory linked with virtual memory + swap file

25. Thin (Over) Provisioning complicates capacity management
Often used when allocating capacity from a pool
Rather than allocating a value, allocate a minimum and maximum value
The min value is reserved and guaranteed (exclusive allocation)
The capacity from min to max is not guaranteed (shared allocation)
The maximum value can be used to determine the unallocated capacity and also to determine the amount of overprovisioning – the pool may limit the total allocated or use some other algorithm (e.g. allowing 20% overprovisioning)

26. Capacity Allocation Host VM 1 Providing Consuming Max Allocation CPU
Overhead
Unallocated
Allocated
Allocated
Guaranteed Allocation
Max Allocation
CPU
VM 2
Overhead
Unallocated
Allocated
Memory
Max Allocation
Allocated
Guaranteed Allocation

27. Resource Capacity Provision
Total
Overhead
Unallocated
Allocated
Actual Usage (highly variable, from 0 to Max Allocation
Shared Allocation
Allocatable
Exclusive Allocation
Note that some of these values can be derived from the others, so not all need to be stored in the model

28. Resource Capacity Allocation
Max Allocation (limit, may have no limit)
Allocated
Shared Allocation
Actual Usage (highly variable, from 0 to Max Allocation
Total consumed
Exclusive # Allocation
Overhead*
Note that some of these values can be derived from the others, so not all need to be stored in the model
# = Guaranteed / Reserved Allocation

29. Capacity Clustering & Pooling
Host Cluster
Resource Pool 1
Provision VM
Host 1
Need to understand how to aggregate and partition allocations
Pooled Provision
Allocation
Provision
Allocation
Overhead
Unallocated
Allocated
Overhead
Unallocated
Allocated
Allocated
Allocated
Overhead
Unallocated
Allocated VM
Host 2
Overhead
Unallocated
Allocated
Allocated
To another Pool