Open Stack

From: Ricardo Aravena [mailto:raravena80 at gmail.com]
Sent: Friday, February 23, 2018 10:32 AM
To: Tao Peng <bergwolf at hyper.sh>
Cc: Larrew, Jesse <Jesse.Larrew at amd.com>; Hollingsworth, Brent <brent.hollingsworth at amd.com>; Woller, Thomas <thomas.woller at amd.com>; Kaplan, David <David.Kaplan at amd.com>; kata-dev at lists.katacontainers.io
Subject: Re: [kata-dev] Kata with AMD Secure Encrypted Virtualization (SEV)



On Fri, Feb 23, 2018 at 5:55 AM, Tao Peng <bergwolf at hyper.sh<mailto:bergwolf at hyper.sh>> wrote:
On Fri, Feb 23, 2018 at 12:32 AM, Larrew, Jesse <Jesse.Larrew at amd.com<mailto:Jesse.Larrew at amd.com>> wrote:
> Ah! Yes, live migration of encrypted guests is supported by the hardware. Qemu support is still being actively developed though. The details of migrating encrypted VMs are discussed in our Secure Encrypted Virtualization API [1]. Basically, the AMD Secure Processor re-encrypts and integrity-protects the guest memory into discrete "packets" that can be sent to the destination machine. The transport/integrity keys used for migration are ephemeral keys negotiated with the receiving machine using a Diffie-Hellman exchange.
>
> Live migration is useful for VMs that need to be stateful. However, my impression of container use cases is that they are encouraged to be stateless. For stateless containers, it would most likely be quicker to simply throw away the container and start a new one on the destination machine. Do you have specific use cases that require the ability to migrate containers?
>
Yes. We have an optimization [1] based on the qemu live migration
feature which let us share the initial part of guest memory among
guests on the same host. It's a quite useful feature for vm-based
container workload because the kernel and initramfs are most likely
the same for all guests on the host.

My two cents here:

It's also quite useful at a higher level when you are dealing with stateful services, that are not heavily containerized in prod yet. First ones that pop in my head are redis/memchached. Then there's also big data stateful type of dbs that require a lot memory, say Cassandra, or Scylla.

That’s good information. Thanks! I’m definitely trying to understand which workloads would benefit most from the container-in-a-VM approach, and which features/capabilities are most important.

Are there any other features/tricks that an SEV implementation would need to preserve?

Sincerely,
Jesse
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.katacontainers.io/pipermail/kata-dev/attachments/20180223/5465c760/attachment.html>