Re: [Xen-devel] Xen / EC2 release criteria proposal

[Matt Wilson <msw@xxxxxxxx>]

On Fri, Aug 09, 2019 at 05:56:11PM -0700, Adam Williamson wrote:
[...]
> So it seems like this would also be a good opportunity to revisit and
> nail down more specifically exactly what our cloud requirements are.
> bcotton suggested  that we require two sample instance types to be
> tested, c5.large (KVM) and t3.large (Xen). (I've also mailed Thomas
> Cameron, ex-of Red Hat, now of Amazon, for his opinion, as it seemed
> like it might be worthwhile - he's promised to get back to me).
> 
> So, for now, let me propose this as a trial balloon: we rewrite the
> above criterion to say:
> 
> "Release-blocking cloud disk images must be published to Amazon EC2 as
> AMIs, and these must boot successfully and meet other relevant release
> criteria on c5.large and t3.large instance types."

Hi Adam,

Thanks for bringing this up. It's good to revisit things from time to
time as the world changes.

Of the two instances that you propose, neither runs on Xen. The T2
instances run on Xen, but T3 uses the KVM-based Nitro hypervisor.

To ensure that a Linux based AMI functions across all of the devices
and operating modes of EC2, you need to cover:

x86 platforms
-------------
* Xen domU with only PV interfaces (e.g., M3 instances)
* Xen domU with Intel 82599 virtual functions for Enhanced Networking
  (e.g., C3 instances running in a VPC)
* Xen domU with Enhanced Networking Adapter (e.g., R4 instances)
* Xen domU with NVMe local instance storage (e.g., virtualized I3
  instances)
* Xen domU with more than 32 vCPUs (e.g., c4.8xlarge)
* Xen domU with four NUMA nodes (e.g., x1.32xlarge)
* Xen domU with maximum RAM available in EC2 (x1e.32xlarge)
* KVM guest with consistent performance (e.g., c5.large)
* KVM guest with burstable performance (e.g., t3.large)
* KVM guest with local NVMe storage (e.g., c5d.large)
* KVM guest with 100 Gbps networking and Elastic Fabric Adapter
  (c5n.18xlarge)
* KVM guest on AMD processors (e.g., m5a.large)
* KVM guest on AMD processors with maximum NUMA nodes (e.g.,
  m5a.24xlarge)
* Bare metal Broadwell (i3.metal)
* Bare metal Skylake (m5.metal)
* Bare metal Cascade Lake (c5.metal)

Arm platforms
-------------
* KVM guest on Arm with 1 CPU cluster (a1.xlarge)
* KVM guest on Arm with 2 CPU clusters (a1.2xlarge)
* KVM guest on Arm with 4 CPU clusters (a1.4xlarge)

Not all of these are going to cause an image to fail to boot up to the
point where a customer can SSH in. But a good number of these have
caused early boot problems in the past (e.g., >32 vCPUs on PVHVM Xen).

--msw

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel