Skip to content

On DPDK and RDMA Related Software

Version History
Date Description
Feb 16, 2021 Some updates on Mellanox RDMA NICs
Dec 14, 2020 More on DPDK
May 28, 2020 Copied from summary

This note mainly talks about how DPDK interacts with RDMA (libibverbs), and how libibverbs communicates with the kernel. I document some misc things about RDMA as well.

RDMA NIC Latest Updates

I sometimes read the MLNX_OFED to track the latest changes introduced in RDMA NICs. They are not sorted chronologically.

  1. Advanced Transport
    • This section talks about XRC and Dynamically Connected Transport (DCT). Those are not new, they have been around for some time. I’m really not sure whether anyone is using them.
    • The RDMA scalability issue stems from the stateful RDMA QP/MR and limited on-chip SRAM cache. Many prior work tried to address them. The latest work in this space are: FLOCK, SOSP‘21 that multiplex QPs in SW; LITE, SOSP‘17; FaRM/FaSST/etc.
  2. Mellanox Zero Touch RoCE.
    • Came across a thing called Zero Touch RoCE, looks like it essentially is RoCE w/o PFC.
    • Based on the description, ConnectX-6 is actually using Selective Transmission to handle lossy RoCE!
    • Wow, a lot of changes made to the CC algorithm. Apparently, they must be.
    • So in all, they changed the retranmission mechanism and CC (the whole transport) to make RDMA NIC work with lossy links (i.e., no PFC). This seems a milestone to me.
  3. Out-of-Order (OOO) Data Placement
    • Interesting. So they now will not drop out-of-sequence/order packets. This of course is not their original Go-Back-N retranmission protocol, but this mechanism works well with data center multi-path routing (e.g., ECMP) and helps improve network utilization.
    • Looks like that this technique, along with the above Zero Touch RoCE, essentially transforms the original Go-Back-N based RDMA transport that best to work with PFC, into one that is Selective Retransmission-based and can work w/o lossless link layer.
    • This is of course not impossible and not difficult. In their OOO placement scheme, they can directly move OoO packets into host DRAM without even caching them in on-chip memory/cache (not possible!). So the cost is really minimal, maybe a set of bitmaps. They probably use techniques in the IRN, SIGCOMM‘19 paper to track the not-fully-received msgs.
    • The end result is nice. The RDMA NIC can now get rid of its reliance on lossless link layer (IB or PFC-based Ethernet). So many PFC issues can be avoided if you are using RoCE. Just like the IRN paper mentioned, eventually, the iWRAP choice wins.
  4. Device Memory Programming
    • (Thank you Stew for pointing me to this feature. It is used in the Sherman, SIGMOD‘22 paper)
    • The RDMA NIC on-device memory is exposed to user applications. RDMA verbs can directly access them. This avoids the PCIe trips to main memory. Great performance indeed. But I’m not sure how large it is and how to properly manage it.
    • Do note that it is quite easy for any FPGA-based SmartNICs to have this sort of feature implemented.


DPDK leverages VFIO to be able to directly access physical devices in the user space. Note that QEMU/Firecracker also use VFIO to directly assign devices to guest OSes (i.e., device passthrough mode).

Although both DPDK and RDMA’s data path bypass kernel, their control path are very different from each other. For most NIC drivers in DPDK, there are completely self-contained device drivers in the user space, and these drivers can directly communicate with the hardware device via MMIO (all possible thanks to VFIO). Specifically, once DPDK has done some VFIO ioctls, all data and control path can bypass kernel. Nice, right´╝č

However, for the rdma-core, a lot of the control-path IB verbs (e.g., create_pd, create_cq) still communicate with the kernel via ioctl calls on Infiniband related device files. On the kernel side, the in-kernel uverb hanlders are located in drivers/infiniband/core/uverbs.c. Do note that this is a quite complicated way to build communicatation channels between user and kernel space, although it is quite efficient. This simple framework is used by several other kernel subsystems, such as io_uring. In details, the control verbs mmap some pages between user and kernel, then all the following data path IB verbs (e.g., post_send) could just bypass kernel and talk to the device via MMIO directly. Though rdma-core also has some vendor-specific “drivers”, this is really different from the above DPDK’s userspace PCIe driver. Userspace “rdma-core” vendor-driver deals with the kernel devel vendor-level driver details (same for the ones inside DPDK).

FWIW, if you are using a Mellanox VPI card in Ethernet mode (e.g. CX3-5), DPDK will use its built-in mlx driver, which further use libibverbs, which further relies on kernel IB stack. It’s not a complete user solution somehow. Note that DPDK built-in mlx driver uses RAW_PACKET QPs.

  • image

DPDK Internal


  • The user-facing part is called Envionmemt Abstraction Layer (EAL), which provides a set of portable interfaces among many OSes. We can think it of as a “POSIX” interface. This EAL has quite a lot useful and handy APIs, e.g., multicore support where you can call a function on arbitray cores (like the linux on_each_cpu core), timers, atomic operations, memory management APIs. I have built all these components myself, still very pleased to see this.
  • Poll Mode Driver - we cover the mlx ones above
  • Various other drivers


Below is a list of RDMA-based systems I have used or the ones I think are useful.

For RDMA programming tricks, see this seminal work: Design Guidelines for High Performance RDMA Systems, ATC‘16

  • Mellanox libvma
    • An userspace IB verbs based layer providing POSIX socket APIs. (The SocketDirect, SIGCOMM‘19 paper was building a similar thing).
  • verbs perftest
    • The collection contains a set of bandwidth and latency benchmark such as:
    • Send - ib_send_bw and ib_send_lat
    • RDMA Read - ib_read_bw and ib_read_lat
    • RDMA Write - ib_write_bw and ib_wriet_lat
    • RDMA Atomic - ib_atomic_bw and ib_atomic_lat
    • Native Ethernet (when working with MOFED2) - raw_ethernet_bw, raw_ethernet_lat
  • rdma-core
    • This is the core userspace IB verbs library (e.g., libibverbs). Whenever you are writing userspace RDMA applications, you are using this library.
    • It is interesting to learn how userspace IB layer communicates with kernel. It is using ioctl() and mmap() to do the trick, quite standard. Not sure how io_uring would help here. The ABI interface (i.e., data structures) are quite complex and has several versions.
    • libibverbs/example
      • asyncwatch.c
      • device_list.c
      • devinfo.c
      • pingpong.c
      • rc_pingpong.c
      • srq_pingpong.c
      • uc_pingpong.c
      • ud_pingpong.c
      • xsrq_pingpong.c
    • infiniband-diags
      • ibv_devinfo
      • iblinkinfo
      • ibping
      • ibaddr
    • Kernel Infiniband stack
  • RPC

Last update: February 17, 2022


Back to top