This document is for a development version of Ceph.

Ceph messenger DPDKStack

Compiling DPDKStack

Ceph dpdkstack is not compiled by default. Therefore, you need to recompile and enable the DPDKstack component. Optionally install dpdk-devel or dpdk-dev on distros with precompiled DPDK packages, and compile -DWITH_DPDK=ON

Setting the DPDK Network Adapter

Most mainstream NICs support SR-IOV and can be virtualized into multiple VF NICs. Each OSD uses some dedicated NICs through DPDK. The mon, mgr and client use the PF NICs through the POSIX protocol stack.

Load the driver on which DPDK depends:

modprobe vfio
modprobe vfio_pci

Configure Hugepage by editing /etc/sysctl.conf

vm.nr_hugepages = xxx

Configure the number of VFs based on the number of OSDs:

echo $numvfs > /sys/class/net/$port/device/sriov_numvfs

Binding NICs to DPDK Applications: -b vfio-pci 0000:xx:yy.z

Configuring OSD DPDKStack

By default, the DPDK RTE initialization process requires the root privileges for accessing various resources in system. To grant the root access to the ceph user:

usermod -G root ceph

The OSD selects the NICs using ms_dpdk_devs_allowlist:

  1. Configure a single NIC.

    ms_dpdk_devs_allowlist=-a 0000:7d:010


  2. Configure the Bond Network Adapter

    ms_dpdk_devs_allowlist=--allow=0000:7d:01.0 --allow=0000:7d:02.6 --vdev=net_bonding0,mode=2,slave=0000:7d:01.0,slave=0000:7d:02.6

DPDK-related configuration items are as follows:


ms_dpdk_debug_allow_loopback = true


Debug and Optimization

Locate faults based on logs and adjust logs to a proper level:


if the log contains a large number of retransmit messages,reduce the value of ms_dpdk_tcp_wmem.

Run the perf dump command to view DPDKStack statistics:

ceph daemon osd.$i perf dump | grep dpdk

if the dpdk_device_receive_nombuf_errors keeps increasing, check whether the throttling exceeds the limit:

ceph daemon osd.$i perf dump | grep throttle-osd_client -A 7 | grep "get_or_fail_fail"
ceph daemon osd.$i perf dump | grep throttle-msgr_dispatch_throttler -A 7 | grep "get_or_fail_fail"

if the throttling exceeds the threshold, increase the throttling threshold or disable the throttling.

Check whether the network adapter is faulty or abnormal.Run the following command to obtain the network adapter status and statistics:

ceph daemon osd.$i show_pmd_stats
ceph daemon osd.$i show_pmd_xstats

Some DPDK versions (eg. dpdk-20.11-3.e18.aarch64) or NIC TSOs are abnormal, try disabling tso:


if VF NICs support multiple queues, more NIC queues can be allocated to a single core to improve performance:


Status and Future Work

Compared with POSIX Stack, in the multi-concurrency test, DPDKStack has the same 4K random write performance, 8K random write performance is improved by 28%, and 1 MB packets are unstable. In the single-latency test,the 4K and 8K random write latency is reduced by 15% (the lower the latency is, the better).

At a high level, our future work plan is:

OSD multiple network support (public network and cluster network) The public and cluster network adapters can be configured.When connecting or listening,the public or cluster network adapters can be selected based on the IP address.During msgr-work initialization,initialize both the public and cluster network adapters and create two DPDKQueuePairs.