Oracle recently announced the launch of its Standard E3 instances based on the 2nd generation of AMD EPYC. The code name for this new offering is AMD Rome.
Note this is currently available in specific regions. Roll-out to the rest of the regions will come shortly.
The new offering provides 128 cores per server (2x EPYC 7742, 64 cores per socket @2.25Ghz) and up to 2TB of memory (DDR4). It is not affected by Meltdown and Foreshadow security vulnerabilities due to the differences in the paging architecture. It brings higher resiliency against speculative execution attacks compared to other processors.
The offering includes both Baremetal and Virtual Machine instances :
This second-generation of AMD processors not only increases the maximum core count, from 32 to 64 cores, it also increases the L3 cache capacity up to 256MB. However, the biggest change comes in the NUMA architecture, where AMD has reduced the number of domains and hops, improving NUMA attributes for general workloads.
If you want to deep dive into AMD ROME EPYC Architecture I highly recommend the following post.
Getting back to our AMD catalog. Looking at the previous AMD shapes in Oracle Cloud (E2 instances) and comparing it with the new ROME shapes there is a difference in the maximum oCPU and Memory capacity per instance, but more interesting, network bandwidth increases. Previous E2 had a 30% penalty in network bandwidth.
These new instances come with a new feature on the platform. The way you configure your instances is now FLEXIBLE. There is no more a fixed instance catalog (at least for the AMD E3 shapes). You can now have FREEDOM when configuring your instances. This is how it looks now:
There is still a relationship between the number of oCPUs, the number of vNICs, and the network bandwidth that your VM can use. Let’s see a couple of examples:
As you can see above the minimum instance provides 1 Gbps bandwidth and 2 vNICs, but what if I want a VM with odd oCPUs? You can now configure it.
What is the maximum configuration of E3 instances?
You can get up to 24 vNICs and 40Gbps bandwidth per VM. Yes, there is a bug in the console, but that gives you a hint of the number of vNICs you can scale in the future 😉
Let’s deploy couple of instances and see how it looks like. We will deploy two virtual machines (E3) with the minimum configuration in two different Availability Domains:
First of all we will look into the CPU properties. We find and EPYC 7742 64-core processor:
If you are curious about the latency between two Availability domains, here are couple of tests using different packet size:
But let’s get into the network bandwidth. For an E3 VM with 1 oCPU we should have 1Gbps network bandwidth…
The theoretical throughput of Gigabit Ethernet with jumbo frames, using TCP is 987Mbps or 123MB/s (without jumbo frame should be around 928 Mbps). Seems we are hitting full capacity…
If you plan to deploy BM you will get up to 100Gbps (2x50Gbps) of overall bandwidth. Awesome, right?
That’s quite impressive and brings some questions around the underlying infrastructure. Is Oracle doing network oversubscription? Remember that Oracle announced a 25Gbps non-oversubscription network. What has changed?
Stay tuned, we will unveil this in the next post!
For more information please visit E3 Compute Release Announcement