On Wed, Jan 24, 2024 at 2:45 PM Rahul Gupta <rahulg...@linux.microsoft.com> wrote: > > From: Rahul Gupta <rahulg...@microsoft.com> > > In continuation to the following email, I am sending this patch. > (https://inbox.dpdk.org/dev/20231110172523.ga17...@microsoft.com/) > > Initialization requires rte_eal_init + rte_pktmbuf_pool_create which > can consume a total time of 500-600 ms: > a) For many devices FLR may take a significant chunk of time > (200-250 ms in our use-case), this FLR is triggered during device > probe in rte_eal_init(). > b) rte_pktmbuf_pool_create() can consume up to 300-350 ms for > applications that require huge memory. > > This cost is incurred on each restart (which happens in our use-case > during binary updates for servicing). > This patch provides an optimization using pthreads that applications > can use and which can save 200-230ms. > > In this patch, rte_eal_init() is refactored into two parts- > a) 1st part is dependent code ie- it’s a perquisite of the FLR and > mempool creation. So this code needs to be executed before any > pthreads. Its named as rte_eal_init_setup() > b) 2nd part of code is independent code ie- it can execute in parallel > to mempool creation in a pthread. Its named as rte_eal_init_async_setup(). > > In existing applications no changes are required unless they wish to leverage > the optimization. > > If the application wants to leverage this optimization, then it needs to call > rte_eal_init_async() (instead of call rte_eal_init()), then it can create a > thread using rte_eal_remote_launch() to schedule a task it would like todo in > parallel rte_eal_init_async_setup(), this task can be a mbuf pool creation > using- rte_pktmbuf_pool_create() > > After this, if next operations require completion of above task, then > user can use rte_eal_init_wait_async_setup_complete(), > or if user wants to just check status of that thread, then use- > rte_eal_init_async_setup_done()
Looking at what this patch does.. I am under the impression all you really need is rte_eal_init without initial probing. Such behavior can probably be achieved with a allowlist set to a non existing device (like for example "-a 0000:00:00.0"), then later, use device hotplug. Some quick note on this patch. - don't expose symbols externally if they are only for internal use in the same library, - current version is 24.03, not 24.01 (wrt comment in version.map), - other OSes are not handled by this patch, please do the work for FreeBSD and Windows, - as a followup of the previous point, please check if we can share code between OSes and, if so, move the shared code to lib/eal/common, - did you test this series with multiprocess? - why should telemetry and other parts of the current rte_eal_init() be left in the second stage of this initialisation pipeline? -- David Marchand
