Feature Discussion: CPU topology detection on FreeBSD
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD. Currently both eal_cpu_core_id() and eal_cpu_socket_id() in rte_eal library always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.
The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is that the return value is an XML formatted object (see https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/). (https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/) I'm not sure how feasible it is to parse XML inside DPDK.
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Oscar
Quoted 9 lines · Oscar Zhao
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD. Currently both
eal_cpu_core_id()andeal_cpu_socket_id()inrte_eallibrary always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is that the return value is an XML formatted object (see https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/). (https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/) I'm not sure how feasible it is to parse XML inside DPDK.
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Oscar
Even adding in the numa node information would be a welcome start.
Quoted 5 lines · Bruce Richardson
Quoted 4 lines · Oscar Zhao
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD. Currently both
eal_cpu_core_id()andeal_cpu_socket_id()inrte_eallibrary always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.
Quoted 8 lines
Quoted 5 lines
The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is that the return value is an XML formatted object (see https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/). (https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/) I'm not sure how feasible it is to parse XML inside DPDK.
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Oscar
Even adding in the numa node information would be a welcome start.
There's little value in it though, because as far as i'm aware we can't allocate memory on specific NUMA nodes in FreeBSD.
Signature · 2 lines
Thanks, Anatoly
Quoted 10 lines · Anatoly
Quoted 9 lines · Bruce Richardson
Quoted 8 lines · Oscar Zhao
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD. Currently both
eal_cpu_core_id()andeal_cpu_socket_id()inrte_eallibrary always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is that the return value is an XML formatted object (see https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/). (https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/) I'm not sure how feasible it is to parse XML inside DPDK.
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
Oscar
Even adding in the numa node information would be a welcome start.
There's little value in it though, because as far as i'm aware we can't allocate memory on specific NUMA nodes in FreeBSD.
That is because the NUMA support in FreeBSD was not in place when we first did the DPDK port. However, I still see no reason not to start adding it now, even if it is only for information purposes at this point.
FreeBSD rolled their own NUMA memory APIs on 12.0 (see https://www.freebsd.org/cgi/man.cgi?query=cpuset_getdomain&sektion=2&apropos=0&manpath=FreeBSD+12.2-RELEASE+and+Ports). These syscalls allow binding of a process/thread/jail memory allocation to a specific NUMA domain.
I think the first step is to support correct NUMA detection. The "de facto" way of doing that is parsing the XML output from the cpu topology sysctl. I'm not sure how viable it is in DPDK without introducing external dependencies. I've also been playing with the hwloc library which supposedly offers portable topology detection.
If both options are not feasible maybe something can be done on the FreeBSD side first.
Quoted 31 lines · Bruce Richardson
Quoted 19 lines · Anatoly
On 29-Jan-21 10:05 AM, Bruce Richardson wrote:
Quoted 3 lines · Oscar Zhao
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD.
Currently both
eal_cpu_core_id()andeal_cpu_socket_id()inrte_eallibrary always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.Quoted 2 lines
The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is
that the return value is an XML formatted object (see
https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299).
(https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299) I'm not sure how feasible it is to parse XML inside DPDK.Quoted 2 lines
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl
dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Quoted 1 line
Quoted 6 lines
Quoted 1 line
Oscar
Even adding in the numa node information would be a welcome start.
There's little value in it though, because as far as i'm aware we can't allocate memory on specific NUMA nodes in FreeBSD.
That is because the NUMA support in FreeBSD was not in place when we first did the DPDK port. However, I still see no reason not to start adding it now, even if it is only for information purposes at this point.
Quoted 16 lines · Bruce Richardson
Quoted 15 lines · Anatoly
Quoted 13 lines · Bruce Richardson
Quoted 9 lines · Oscar Zhao
Hello dpdk devs,
I would like to discuss the feasibility of implementing CPU topology detection on FreeBSD. Currently both
eal_cpu_core_id()andeal_cpu_socket_id()inrte_eallibrary always return 0 on FreeBSD, making NUMA-aware development impossible without resorting to external libraries or system APIs.The CPU topology information is available via sysct kern.sched.topology_spec. The only issue is that the return value is an XML formatted object (see https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/). (https://forums.freebsd.org/threads/number-of-cpus-and-cores.41299/) I'm not sure how feasible it is to parse XML inside DPDK.
On a side note, obtaining the physical NUMA node id of a core is easy - sysctl dev.cpu.[cpu#].%domain returns the corresponding node id but I have yet found a straightforward way to detect hyperthreads besides the XML thing.
Oscar
Even adding in the numa node information would be a welcome start.
Quoted 8 lines
Quoted 3 lines
There's little value in it though, because as far as i'm aware we can't allocate memory on specific NUMA nodes in FreeBSD.
That is because the NUMA support in FreeBSD was not in place when we first did the DPDK port. However, I still see no reason not to start adding it now, even if it is only for information purposes at this point.
A lot of DPDK apps will fail to work in this case, because most of them will allocate with an explicit NUMA node. So, if you have a NUMA node assigned to an lcore but can only allocate from NUMA node 0, any allocation explicitly targeting NUMA node 1 will not succeed.
Signature · 2 lines
Thanks, Anatoly