How this ID would be used? Would it be added as a parameter to many of the
create calls (shm, pool, timer, crypto, …)?
I’d like to avoid introduction of explicit “socket” IDs (if possible).
Application should not care too much on physical format/layout of the circuit
boards or SoCs. In general, the problem is about how *user* could control
clustering and get information on cost of crossing various cluster borders. Do
we need API calls and IDs, or just implementation defined named resources that
user can pass down (e.g. on command line).
For example, an application could run in SMP setup on a blade that have 6 SoCs
connected like this (with 2-3 coherent IO links toward each neighbor). Each SoC
would have clustering inside: e.g. ring based interconnect, 8x8 CPUs in
clusters, 3-4 cache levels, DDR memory controllers and HW accelerators spread
around the interconnect. Also virtualization/VMs could hide some of these
details (which virtual CPU is close to which virtual interface, etc).
DDR DDR
| |
SoC1 SoC2 SoC3
+ ------ + ------ +
| | |
+ ------ + ------ +
SoC4 SoC5 SoC6
| | |
eth0 eth1 DDR
Memory/HW accelerator access latency would vary a lot depending on the location
of the source CPU and target device (memory). E.g. latency from CPU:X/SoC1 to
DDR:Y on SoC2, SoC3 and SoC6 would be quite different. Also it could be hard
to tell without testing, if optimal placement for a HW buffer pool used for
packet input from eth0 and output to eth1 would be SoC4, 5 or 6 (if in addition
application SW runs on SoC6).
-Petri
From: lng-odp [mailto:[email protected]] On Behalf Of EXT Bill
Fischofer
Sent: Saturday, January 30, 2016 12:40 AM
To: Zoltan Kiss
Cc: LNG ODP Mailman List
Subject: Re: [lng-odp] [RFC API-NEXT PATCH 2/2] api: cpu: add routines for
obtaining socket ids
We'll be discussing this RFC on Monday's ARCH call. I'm certainly open to
adding this if it's deemed useful. I do note that DPDK defines SOCK_ID_ANY to
be -1 so I'd like to minimize translation overhead for odp-dpdk in this area as
well.
On Fri, Jan 29, 2016 at 10:27 AM, Zoltan Kiss
<[email protected]<mailto:[email protected]>> wrote:
On 29/01/16 02:44, Bill Fischofer wrote:
Add odp_cpu_socket_id() and odp_cpu_socket_id_cpu() routines
Signed-off-by: Bill Fischofer
<[email protected]<mailto:[email protected]>>
---
include/odp/api/cpu.h | 23 +++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/include/odp/api/cpu.h b/include/odp/api/cpu.h
index 4cbaf58..fe74825 100644
--- a/include/odp/api/cpu.h
+++ b/include/odp/api/cpu.h
@@ -36,6 +36,29 @@ extern "C" {
int odp_cpu_id(void);
/**
+ * CPU socket id
+ *
+ * Returns the socket id associated with the calling CPU on NUMA systems.
+ * Socket ID numbering is system specific.
+ *
+ * @return Socket ID of the calling CPU
+ * @retval ODP_SOCKET_ID_ANY If the caller is not running on a NUMA system.
+ */
+uint32_t odp_cpu_socket_id(void);
+
+/**
+ * CPU socket id of designated CPU
+ *
+ * Returns the socket id associated with a specified CPU on NUMA systems.
+ * Socket ID numbering is system specific.
+ *
+ * @return Socket ID of the designated CPU
+ * @retval ODP_SOCKET_ID_ANY If the specified CPU is unknown or caller is
+ * not running on a NUMA system.
Maybe worth to differentiate between the two scenario with different return
value? To query for an invalid CPU sounds like a serious problem in the
application while the other is part of normal operation.
+ */
+uint32_t odp_cpu_socket_id_cpu(int cpu_id);
+
+/**
* CPU count
*
* Report the number of CPU's available to this ODP program.
_______________________________________________
lng-odp mailing list
[email protected]
https://lists.linaro.org/mailman/listinfo/lng-odp
