In multi node ODP implementation / application usage perceptive, we need to consider, How we can expose the HW resources in each node. resources could be cpus, memory and any hw accelerated blocks for packet processing.
In case of CPU resource, we could take the current API model like, API's for querying how may cpu resource available in each node and start specific work on selected cpus using odp_cpu_mask_t/ Let implementation take care of pinning/exposing the number cores for ODP on each node. In case of memory resource, IMO odp_shm_reserve can extended to allocated form a specific node In case of hw accelerated blocks resources, IMO we should add node parameter while creating the handles IMO, Gábor Sándor Enyedi's example may be visualized like this on multi node ODP -local_pool = odp_pool_create() // create a local pool -odp_pktio_open(..,local_pool) // open local node pktio and attach to local pool -remote_pool = odp_pool_create_onnode(node...) // create a remote pool as packet needs to go remote node DDR -odp_pktio_open_onnode(node,...,remote_pool) // open remote node pktio with remote pool -odp_cpu_count() -create cpu mask and lunch work on local node -odp_cpu_count(node) // to get number works available on remote node -create cpu mask and lunch work on remote node From: Bill Fischofer <bill.fischo...@linaro.org> Sent: Friday, May 8, 2015 7:43 PM To: Gábor Sándor Enyedi Cc: Savolainen, Petri (Nokia - FI/Espoo); Jacob, Jerin; Zoltan Kiss; lng-odp@lists.linaro.org Subject: Re: [lng-odp] NUMA aware memory allocation? Thanks, that's good info. So in this case is it sufficient to say that the memory used for odp_pool_create() is the one associated with the thread that executes the create call? Presumably then when a packet arrives and is assigned to a CoS that points to that pool then events from that pool are sent to queues that are only scheduled to the corresponding cores that have fast access to that pool. Right now queues have an odp_schedule_group_t but that's still fairly rudimentary. It sounds like we might want to point the queue at the pool for scheduling purposes so that it would inherit the NUMA considerations you mention. On Fri, May 8, 2015 at 9:00 AM, Gábor Sándor Enyedi <gabor.sandor.eny...@ericsson.com> wrote: For me and for now the use-case is very simple: we have an x86 with two Xeon CPU-s (dual socket) in it. Each of the CPU-s have its own memory and own PCIExpress bus, as usual. First, I want to make only some test code, but later we may want to port our high speed OF soft switch to ODP (now, its on DPDK). We want to assign a correct core for each interface, and each slot must use its own copy of forwarding data in its own memory. We have the experience that if we accidentally assigned a bad core to an interface, we could get even about 50% performance drop, so NUMA is essential. Based on the previous, for us something similar to that used in DPDK's rte_malloc (and its variants) and a NUMA aware buffer pool create was enough for now. Later we want to investigate other architectures... but I don't know the use-cases yet. Gabor On 05/08/2015 03:35 PM, Bill Fischofer wrote: Insofar as possible, the mechanics of NUMA should be the responsibility of the ODP implementation, rather than the application, since that way the application retains maximum portability. However, from an ODP API perspective, I think we need to be mindful of NUMA considerations to give implementations the necessary "hooks" to properly support the NUMA aspects of their platform. This is why ODP APIs need to be careful about what addressability assumptions they make. If Gábor or Jerrin can list a couple of specific relevant cases I think that will help in focusing the discussion and get us off to a good start. On Fri, May 8, 2015 at 8:26 AM, Savolainen, Petri (Nokia - FI/Espoo) <petri.savolai...@nokia.com> wrote: Hi, ODP is OS agnostic and thus thread management (e.g. thread creation and pinning to physical cores) and NUMA awareness should happen mostly outside of ODP APIs. For example, NUMA could be visible in ODP APIs this way: * Add odp_cpumask_xxx() calls that indicate NUMA dependency between CPUs (just for information) * Add a way to identify groups of threads which frequently share resources (memory and handles) within the group * Give the thread group as a hint (parameter) to various ODP calls that create shared resources. Implementation can use the information to allocate resources "near" to the threads in the group. However, the user is responsible to group the threads and map/pin those into physical CPUs in a way that enables NUMA aware optimizations. -Petri > -----Original Message----- > From: lng-odp [mailto:lng-odp-boun...@lists.linaro.org] On Behalf Of ext > Gábor Sándor Enyedi > Sent: Friday, May 08, 2015 10:48 AM > To: Jerin Jacob; Zoltan Kiss > Cc: lng-odp@lists.linaro.org > Subject: Re: [lng-odp] NUMA aware memory allocation? > > Hi, > > Thanks. So, is the workaround for now to start the threads, and do all > the memory reservation on the thread? And to call odp_shm_reserve() > instead of simple malloc() calls? Can I use multiple buffer pools, one > for each thread or interface? > BR, > > Gabor > > P.s.: Do you know when will this issue in the API be fixed (e.g. in next > release or whatever)? > > On 05/08/2015 09:06 AM, Jerin Jacob wrote: > > On Thu, May 07, 2015 at 05:00:54PM +0100, Zoltan Kiss wrote: > > > >> Hi, > >> > >> I'm not aware of any such interface, but others with more knowledge can > >> comment about it. The ODP-DPDK implementation creates buffer pools on > the > >> NUMA node where the pool create function were actually called. > > current ODP spec is not NUMA aware. We need to have API to support nodes > enumeration and > > explicit node parameter to alloc/free resource from specific node like > odp_shm_reserve_onnode(node, ...) > > and while keeping existing API odp_shm_reserve() allocated on node where > the current code runs > > > > > >> Regards, > >> > >> Zoli > >> > >> On 07/05/15 16:32, Gábor Sándor Enyedi wrote: > >>> Hi! > >>> > >>> I just started to test ODP, trying to write my first application, but > >>> found a problem: if I want to write NUMA aware code, how should I > >>> allocate memory close to a given thread? I mean, I know there is > >>> libnuma, but should I use it? I guess not, but I cannot find memory > >>> allocation functions in ODP. Is there a function similar to > >>> numa_alloc_onnode()? > >>> Thanks, > >>> > >>> Gabor > >>> _______________________________________________ > >>> lng-odp mailing list > >>> lng-odp@lists.linaro.org > >>> https://lists.linaro.org/mailman/listinfo/lng-odp > >> _______________________________________________ > >> lng-odp mailing list > >> lng-odp@lists.linaro.org > >> https://lists.linaro.org/mailman/listinfo/lng-odp > > > _______________________________________________ > lng-odp mailing list > lng-odp@lists.linaro.org > https://lists.linaro.org/mailman/listinfo/lng-odp _______________________________________________ lng-odp mailing list lng-odp@lists.linaro.org https://lists.linaro.org/mailman/listinfo/lng-odp _______________________________________________ lng-odp mailing list lng-odp@lists.linaro.org https://lists.linaro.org/mailman/listinfo/lng-odp
