On 28-Jun-18 8:03 AM, Kumar, Ravi1 wrote:
On 22-Jun-18 5:37 PM, Kumar, Ravi1 wrote:
Hi,
As the memory subsystem in DPDK 18.05 is reworked, it has introduced a problem
for AMD EPYC 2P platforms.
The issue is that DPDK 18.05 only works with up to 4 NUMAs. For AMD EPYC 2P
platforms, DPDK now only works with P0 (NUMA 0-3) and does not work with P1
(NUMA 4-7).
The problem can be fixed by reducing some of the default settings of the memory
subsystem.
To solve this issue:
- We can create our own config file for our integrated 10G NIC,
that is for amd_xgbe PMD. This will make amd_xgbe immune to this problem.
- However, when any other NIC (Intel, Mellanox, Cavium or Broadcom
etc.) is plugged into NUMA 4-7, the problem will still be exposed.
- If we only fix it in "config/common_base", it will cover all
cases.
Our current workaround is:
Edit config file "./config/common_base" and change the following line
CONFIG_RTE_MAX_MEM_MB_PER_TYPE=131072
TO
CONFIG_RTE_MAX_MEM_MB_PER_TYPE=65536
Any better solution for this issue is welcome.
We would appreciate if this issue can be fixed in the next release (18.08) so
the STOCK version of DPDK works on AMD EPYC 2P platforms.
Regards,
Ravi
Hi Ravi,
What is the reason behind this limitation? Is it too much virtual memory being
preallocated?
--
Thanks,
Anatoly
Hi Anatoly,
We believe this is true. By default, too much virtual memory is being
preallocated. The result is it can only support up to 4 NUMAs.
Our workaround is to reduce the amount of preallocated virtual memory by half,
so to support up to 8 NUMAs.
Regards,
Ravi
I assume you see a bunch of failed mmap() calls with ENOMEM?
In general, changing base config that way is an OK change, and it won't
even be an ABI break since this memory is allocated at runtime. I just
want to make sure that we fix the underlying problem, rather than the
symptom.
--
Thanks,
Anatoly
