Bhanuprakash Bodireddy <bhanuprakash.bodire...@intel.com> writes: > Keepalive feature is aimed at achieving Fastpath Service Assurance > in OVS-DPDK deployments. It adds support for monitoring the packet > processing cores(PMD thread cores) by dispatching heartbeats at regular > intervals. Incase of heartbeat misses additional health checks are > enabled on the PMD thread to detect the failure and the same shall be > reported to higher level fault management systems/frameworks. > > The implementation uses OVSDB for reporting the datapath status and the > health of the PMD threads. Any external monitoring application can read > the status from OVSDB at regular intervals (or) subscribe to the updates > in OVSDB so that they get notified when the changes happen on OVSDB. > > POSIX shared memory object is created and initialized for storing the > status of the PMD threads. This is initialized by main thread(vswitchd) > as part of init process and will be periodically updated by 'keepalive' > thread. keepalive feature can be enabled through below OVSDB settings. > > enable-keepalive=true > - Keepalive feature is disabled by default. > > keepalive-interval="5000" > - Timer interval in milliseconds for monitoring the packet > processing cores. > > keepalive-shm-name="/ovs_keepalive_shm_name" > - Shared memory block name where the events shall be updated. > > When KA is enabled, 'ovs-keepalive' thread shall be spawned that wakes > up at regular intervals to update the timestamp and status of pmd cores > in shared memory region. This information shall be read by vswitchd thread > and write the status in to 'keepalive' column of Open_vSwitch table in OVSDB. > > An external monitoring framework like collectd with ovs events support > can read (or) subscribe to the datapath status changes in ovsdb. When the > state > is updated, the collectd shall be notified and will eventually relay the > status > to ceilometer service running in the controller. Below is the high level > overview of deployment model. > > Compute Node Controller Compute Node > > Collectd <----------> Ceilometer <--------> Collectd > > OvS DPDK OvS DPDK > > +-----+ > | VM | > +--+--+ > \---+---/ > | > +--+---+ +------------+----------+ +------+-------+ > | OVS |-----> | ovsevents plugin | --> | collectd | > +--+---+ +------------+----------+ +------+-------+ > > +------+-----+ +---------------+------------+ | > | Ceilometer | <-- | collectd ceilometer plugin | <--- > +------+-----+ +---------------+------------+ > > Performance impact > ------------------ > No noticeable performance or latency impact is observed with > KA feature enabled. > > Bhanuprakash Bodireddy (21): > > [10] Patches help update OVSDB with keepalive status > > vswitch.xml: Add keepalive support. > ovsschema: Introduce 'keepalive' column in Open_vSwitch. > dpdk: Add helper functions for DPDK datapath keepalive. > process: Retrieve process status. > Keepalive: Add initial keepalive support. > bridge: Invoke keepalive framework. > keepalive: Add more helper functions to KA framework. > dpif-netdev: Register packet processing cores to KA framework. > dpif-netdev: Dispatch heartbeats for DPDK datapath. > keepalive: Retrieve PMD status periodically. > bridge: Update keepalive status in ovsdb > > keepalive: Add support to query keepalive statistics. > keepalive: Add support to query keepalive status. > dpif-netdev: Add helper function to check false positives. > > [5] Following patches add additional health checks in case of heartbeat > failure. The following can still be improved and WIP. > > dpif-netdev: Add additional datapath health checks. > keepalive: Check the link status as part of PMD health checks. > keepalive: Check the packet statisitcs as part of PMD health checks. > keepalive: Check the PMD cycle stats as part of PMD health checks. > netdev-dpdk: Enable PMD health checks on heartbeat failure. > > keepalive: Display extended Keepalive status. > Documentation: Update DPDK doc with Keepalive feature. >
Hi Bhanu, I've been playing with this a little bit; is it too late to consider tracking 'threads' instead of 'cores'? I'm not sure what it means for a particular core ID to be 'healthy' - but I know what 'pmd24' not responding means. Additionally, I'd suggest keeping words like 'healthy', and 'unhealthy' out of it. I'd basically just have this keepalive report things on the thread you *know* - last time it poked your status register (and you can also track things like cpu utilization, etc, if you'd like). Then let your higher level thing that reads ceilometer make those "healthy" determinations. After all, sometimes 0% utilization is "healthy," and sometimes it isn't. Just food for thought. -Aaron _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
