
VEXXHOST at OpenInfra Day North America 2026
VEXXHOST recaps OpenInfra Day North America 2026 in Chicago, highlighting its headline sponsorship, Mohammed Naser's talks, and the future of OpenStack operations.
Read field noteField notes / Latest
Engineering notes from operating open infrastructure: the failures, design decisions, and upstream work that make open infrastructure better.
Browse all field notes
VEXXHOST recaps OpenInfra Day North America 2026 in Chicago, highlighting its headline sponsorship, Mohammed Naser's talks, and the future of OpenStack operations.
Read field note
Ceph vs SAN: Compare performance, scalability, licensing, and operational costs to choose the right enterprise storage modernization strategy.
Read field note
Cloud exit strategy checklist: Review data ownership, exit costs, vendor lock-in, and migration support before signing a cloud contract.
Read field noteSR-IOV vs DPDK for packet-heavy VMs, compared on host cores, patch windows, security group coverage, NIC vendor lock-in and staffing.
Virtual network functions such as telecom packet cores, virtual firewalls and load balancers spend their working life moving packets. On the standard kernel virtual switch path, these VMs can need more packet throughput than the host gives them. That is when a faster path goes up for approval.
There are two common answers:
This comparison sets SR-IOV ports assigned directly to VMs against Open vSwitch running DPDK on the host. Two designs sit between them: hardware offload, which uses SR-IOV virtual cards while the virtual switch keeps control of the traffic, and DPDK running inside a VM on an SR-IOV port.
The choice is a design and budget decision as much as a networking one. Each path moves cost to a different line. Pick wrong and you can:
OpenStack's 2025.1 release, on April 2, 2025, added true live migration for SR-IOV devices passed through by flavor on newer kernel drivers. Direct-mode SR-IOV network ports attached through Neutron were left on the older path, which drops the connection during a move. That gap still holds in the 2026.1 release. Each packet-heavy workload needs its own answer: SR-IOV, DPDK, or the standard path it already has.
DPDK's cost is capacity. Open vSwitch runs its DPDK packet processing on polling threads pinned to specific host cores, and its own documentation describes those threads as CPU bound and best run on isolated cores. Those cores run the switch around the clock.
For a budget, that means:
SR-IOV spends no host cores on switching, because the card does that work in hardware. Its costs land elsewhere, in the constraints covered in the next three sections.
In budget terms, DPDK costs a known number of cores on every host in the pool, and SR-IOV costs flexibility on the hosts that run it. On a few large hosts carrying network functions, the reserved cores are a small share of each host. On a dense fleet of general VMs, those cores are capacity you planned to sell.
Live migration decides whether a host can be patched, rebooted or retired without asking a workload owner for a maintenance window. The two paths differ most here.
OpenStack has been able to live-migrate VMs with SR-IOV ports since its Train release in October 2019, with conditions. For SR-IOV ports in direct mode, the virtual card is detached before the move and re-attached afterwards, so the VM loses that connection while its host is drained. A VM can avoid the gap only if its operating system bonds the SR-IOV port with a second port that migrates cleanly: a standard port or an indirect-mode SR-IOV port. OpenStack 2025.1 added a newer migration path for SR-IOV devices passed through by flavor, such as some GPUs, on recent kernel drivers. Direct-mode SR-IOV network ports attached through Neutron do not use it: through the 2026.1 release they are still detached and re-attached.
For planning:
DPDK ports are virtual switch ports. Once the destination host has the same DPDK setup, VMs on them go through OpenStack's normal live migration, with the brief pause any live migration has. Test it with your workloads before you build a patch plan on it. The virtual switch itself restarts during platform upgrades, so ask any provider how long the DPDK data path is down during that restart.
For a regulated workload, the security-control questions in this section usually settle the decision. Confirm with your auditors which security control they accept before you choose.
Security groups are the cloud's built-in firewall rules on each VM port. Neutron's documentation states that security groups are not supported when using SR-IOV, and the firewall driver on those ports must be disabled. SR-IOV traffic goes from the guest to the card without passing the virtual switch where the rules live. With DPDK, traffic still crosses the virtual switch, so Neutron can still apply security groups: as flows of the native Open vSwitch firewall driver with ML2/OVS, or as OVN access control lists with ML2/OVN.
If an audit control says every VM port is filtered by cloud security groups, each SR-IOV port needs a documented compensating control:
That third option makes a virtual firewall appliance on SR-IOV a defensible design. For a general application VM, the compensating control has to be built, documented and audited on its own, and that cost belongs in the business case.
Both paths need specific hardware. The vendor dependency sits in a different layer for each, and that decides who carries the risk when you change cards.
A team that runs one server model and one card can absorb either. Both paths add work for each new NIC model. With SR-IOV, that work also reaches the guest images.
Our hosted and on-premise editions of Atmosphere, the platform we build for running OpenStack, support custom hardware and advanced tuning under fully managed operations.
The strongest case for SR-IOV holds. It avoids the dedicated core budget a DPDK switch needs and puts no software switch in the packet path. A network function built to fail over at the application level also loses little when its host is drained. For a virtual firewall or a telecom function with its own redundancy, the migration and security group arguments above lose most of their weight.
The case for DPDK weakens when the hosts are small, since a fixed core reservation is a larger share of a small host. It also weakens when a single workload needs more packets per second than a few dedicated cores can switch.
No NIC purchase removes the patch-window cost today, because OpenStack through its 2026.1 release still detaches direct-mode SR-IOV network ports during live migration. Two things would move more workloads toward SR-IOV:
Bring these figures to the decision:
Then apply the rule in this order:
Our own engineering has gone into the DPDK side. Since October 2025, Atmosphere's Open vSwitch images are built with AVX-512 for compatible Intel CPUs. Our ovsinit work restarts the virtual switch in place, which cuts DPDK data path downtime during upgrades to about 3 seconds in our testing. Atmosphere supports both SR-IOV and DPDK, and its public deployment code documents DPDK and SR-IOV-based hardware offload for your engineers to review.
Our OpenStack work also reaches past our own platform. Since 2016, our infrastructure has powered much of the OpenDev CI system, where the OpenStack community develops and tests its software.
Bring your workload measurements, hardware plans and maintenance requirements, and our team can work through these trade-offs with yours for each packet-heavy workload, so the rule above becomes a budget line.
Choose from Atmosphere Cloud, Hosted, or On-Premise.
Simplify your cloud operations with our intuitive dashboard.
Run it yourself, tap our expert support, or opt for full remote operations.
Leverage Terraform, Ansible or APIs directly powered by OpenStack & Kubernetes