Wednesday at VMworld. Full out sessions, last day of the Solutions Exchange. Let's check in.
Swing and a Miss
I had a breakout session booked early in the morning this morning. Early enough that I had to skip breakfast to get to it. No problem, I thought, the topic looks relevant to some things going on at work, it'll be worth it. When I originally scheduled the session, it had a vendor-neutral title and description. Twenty minutes into the session, and it was obvious that the speakers sole purpose was to pitch his vendor's products. Not only that, it was further clear that his audience was (potential) partner companies, not customers. Ooof.
That's something you end up learning with time at VMworld. While the vast majority of sessions are well delivered, with interesting, relevant topics, every once in a while you get a pure sales pitch. Know when to cut your loses. That's what I did, so let's move on.
The Virtually Speaking Podcast LIVE: The Future of IT - How Will You Fit in? [HCI1894PU]
As a fan of the Virtually Speaking Podcast, I was looking forward to this session. It was interesting and fun seeing Pete Flecha and John Nicholson basically run through what they would normally do when recording their podcast. It was further helped out by some great guest. Namely Ken Werneburg, Duncan Epping, Frank Denneman, and William Lam.
If you can, it's worth listening to this podcast when it's released. In the meantime, here are two book recommendations that were made during the chat:
Made to Stick: Why Some Ideas Survive and Others Die by Chip Heath
Confessions of a Public Speaker by Scott Berkun
60 Minutes of Non-Uniform Memory Architecture [HBI2278BU] (Redux!)
Following the power outage from the previous day, Frank Denneman's session was rescheduled for Wednesday. It was well worth shuffling my schedule around to attend.
Here are just a few notes I took. The session was very detailed, so I spent more time paying attention than jotting down notes.
When a memory module is under high load, the clock speed is throttled to compensate. On average the memory module's performance goes down 22%.
Node interleaving creates a single logical NUMA node. This increases latency on average compared to local NUMA node usage.
The latest versions of ESXi are designed to take advantage of AMD's EPYC v2 CPUs.
If a process is scheduled on hyper-threading, it is "given back" some CPU scheduling time. This allows it to have more CPU time to compensate for HT not being a full core.
Set numa.preferHT if you have workloads that have high memory requirements but low CPU requirements.
- This helps address memory latency, and can be set per host or per VM.
With CPU hot add, vNUMA is disabled.
You can continue to add vCPU to the VM, but there will be no optimization done.
Hot adding RAM has no effect on vNUMA.
The vNUMA state for the VM is stored in the VMX file.
The "fix" for a VM with CPU hot add enabled, and addtional vCPUs added, is to reboot the VM and reconfigure it so that vNUMA is enabled (disable hot add CPU).
When this session's video is made available online, I highly recommend watching it to get the full story on NUMA.
VMworld Fest
The VMworld Fest was well done this year. It was kicked off by surprise act Billy Idol (although it might have been the worst kept secret at the conference), and followed up by One Republic. Everyone, including myself, enjoyed the conference party this year. It's almost as if, a few years ago in Las Vegas, VMworld noticed that people were skipping the VMworld Fest to go see Billy who was in town that week. Hmm.
In any event, it was well put together and much appreciated.
One Last Sleep
That does it for the penultimate day of VMworld US 2019. Thursday is the final day of the conference and, as hinted above, I'll be spending it at the Future:NET sister conference. Stick around to find out how that goes.