Repository navigation
Changes since v2026.07.00 for release v2026.11.00 (package version 2026.11.0).
DAQIRI 2026.11 imrpoves the ibverbs engine performance for the raw ethernet traffic, adds direct polling mode, adds hardware packet reorder, and adds support hardware timing for tx and rx. Applications can manage queues and memory at runtime and control how packets are placed in CPU or GPU memory. New configuration tools and a TensorRT example help developers build and test data-acquisition pipelines.
Highlights
- ibverbs performance improvements. Smaller packet buffers, batched transmission, and more efficient queue access reduce processing overhead. ibverbs now achieves higher or similar performance to dpdk across all packet sizes and is the default engine for raw ethernet traffic.
- Direct polling mode. Applications can poll raw ibverbs queues and RDMA connections directly, avoiding the handoff to a worker thread, allowing for low-latency rx and tx tailored to time-sensitive packet processing. Direct polling is optional; indirect polling remains the default.
- Hardware packet reorder. Raw ibverbs can use the network adapter to place packet data directly into ordered CPU or GPU buffers, reducing the work needed to reorder it in software. This optional path supports repeating packet sequences on ConnectX-7 or newer; software reorder remains the default.
- Hardware timestamps and scheduling. Applications can record receive times using the network adapter's clock and schedule operations for a chosen time.
- Expanded RDMA operations. C++ and Python applications can read from and write to memory on another system, including writes that notify the receiving application. Reads require reliable connections.
- Manage resources at runtime. Add or remove raw ibverbs queues and memory regions after initialization, including memory allocated by the application. Named UDP destinations let applications send data without rebuilding packet headers for each transfer; DAQIRI stores and reuses the headers.
- Flexible packet buffers. Raw ibverbs can receive into, or transmit from, up to four CPU/GPU memory regions per packet. Applications can discard unwanted parts of received packets without writing them to memory. With shared transmit buffers enabled, the same stored data can be sent with different headers, avoiding a separate payload copy for each send.
Other improvements, fixes, and documentation
- Deployment options. Shared and static DAQIRI builds can load RDMA libraries at runtime. A PyTorch-based SDK container build definition brings together C++/CUDA development tools, Python bindings, benchmarks, and tuning utilities.
- Flow steering and diagnostics. Raw DPDK and ibverbs can distribute incoming traffic across queues while keeping each flow on one queue. Additional options include MAC-address filtering, ibverbs hardware loopback for testing, and ibverbs packet pacing. Pause-frame diagnostics help identify when Ethernet flow control is slowing traffic.
- Reliability fixes. TCP and UDP receive queues now have size limits, and initialization and shutdown clean up resources more reliably. Fixes also address packet and buffer sizing, CUDA context handling, and GPU selection on IGX Thor.
- Performance reports and documentation. Updated reports cover DGX Spark and IGX Thor performance, including test setups and latency measurements. Refreshed getting-started and configuration guides include explanations of packet layout, flow steering, and packet reorder.
- Configuration tools and onboarding. Generate repeatable configurations from profiles and use
daqiri_config_validateto check them before initializing hardware. Thedaqiri-onboardingskill guides setup, configuration, first benchmarks, and troubleshooting. - TensorRT inference example. The optional ResNet-50 application receives images over raw Ethernet, reorders and converts them on the GPU, then runs TensorRT feature extraction. A walkthrough covers operation across two hosts.
Upgrading from 2026.07
- Check the raw engine. Raw Ethernet streams now use ibverbs when it is included in the build. Set
engine: dpdkexplicitly to retain DPDK behavior, including software loopback. - Rebuild applications and bindings. The default library binary interface (ABI) has changed from 0 to 3. Rebuild applications and Python bindings against the matching SDK. If RDMA libraries load at runtime, deploy matching glue libraries too.
- Check Linux and hugepages. Linux kernel 5.15 or newer is required. A
kind: hugeallocation requires reserved hugetlb pages and fails if they are unavailable. - Check hardware and clocks. Packet pacing, accurate send scheduling, and hardware reorder require ConnectX-7 or newer. Synchronize the network adapter and host clocks with PTP before scheduling sends or comparing their timestamps.
Full changelog: v2026.07.00...v2026.11.00