[dpdk-dev] long initialization time for EAL init

[Balaji K N]

Hello,

I tried by disabling mlockall and not do see long time for EAL init and is done quickly. Thanks for help.

ovs-vswitchd unix:/var/run/openvswitch/db.sock -vconsole:emer -vsyslog:err -vfile:info --no-chdir --log-file=/var/log/openvswitch/ovs-vswitchd.log --pidfile=/var/run/openvswitch/ovs-vswitchd.pid --detach

OVS logs
2020-07-29T09:48:00.233Z|00008|dpdk|INFO|Using DPDK 18.11.6
2020-07-29T09:48:00.233Z|00009|dpdk|INFO|DPDK Enabled - initializing...
2020-07-29T09:48:00.233Z|00010|dpdk|INFO|No vhost-sock-dir provided - defaulting to /var/run/openvswitch
2020-07-29T09:48:00.233Z|00011|dpdk|INFO|IOMMU support for vhost-user-client disabled.
2020-07-29T09:48:00.233Z|00012|dpdk|INFO|POSTCOPY support for vhost-user-client disabled.
2020-07-29T09:48:00.233Z|00013|dpdk|INFO|Per port memory for DPDK devices disabled.
2020-07-29T09:48:00.233Z|00014|dpdk|INFO|EAL ARGS: ovs-vswitchd -w 0000:04:00.0,support-multi-driver=1 -w 0000:42:00.0,support-multi-driver=1 -c 0x1 --huge-dir /mnt/huge_ovs_2M --socket-mem 2048,2048 --socket-limit 2048,2048.
2020-07-29T09:48:00.237Z|00015|dpdk|INFO|EAL: Detected 40 lcore(s)
2020-07-29T09:48:00.237Z|00016|dpdk|INFO|EAL: Detected 2 NUMA nodes
2020-07-29T09:48:00.237Z|00017|dpdk|INFO|EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
2020-07-29T09:48:00.280Z|00018|dpdk|INFO|EAL: 99 hugepages of size 1073741824 reserved, but no mounted hugetlbfs found for that size
2020-07-29T09:48:00.281Z|00019|dpdk|INFO|EAL: Probing VFIO support...
2020-07-29T09:48:00.281Z|00020|dpdk|INFO|EAL: VFIO support initialized
2020-07-29T09:48:01.696Z|00021|dpdk|INFO|EAL: PCI device 0000:04:00.0 on NUMA socket 0
2020-07-29T09:48:01.696Z|00022|dpdk|INFO|EAL:   probe driver: 8086:154d net_ixgbe
2020-07-29T09:48:01.696Z|00023|dpdk|INFO|EAL:   using IOMMU type 1 (Type 1)
2020-07-29T09:48:01.940Z|00024|dpdk|INFO|EAL: Ignore mapping IO port bar(2)
2020-07-29T09:48:02.285Z|00025|dpdk|INFO|EAL: PCI device 0000:42:00.0 on NUMA socket 1
2020-07-29T09:48:02.285Z|00026|dpdk|INFO|EAL:   probe driver: 8086:154d net_ixgbe
2020-07-29T09:48:02.412Z|00027|dpdk|INFO|EAL: Ignore mapping IO port bar(2)
2020-07-29T09:48:02.757Z|00028|dpdk|INFO|DPDK Enabled - initialized

Regards,
Balaji

-----Original Message-----
From: Burakov, Anatoly <anatoly.burakov at intel.com> 
Sent: Wednesday, July 29, 2020 2:38 PM
To: David Marchand <david.marchand at redhat.com>; Balaji K N <balaji.k.n at ericsson.com>
Cc: dev at dpdk.org
Subject: Re: [dpdk-dev] long initialization time for EAL init

On 29-Jul-20 9:52 AM, David Marchand wrote:
> On Wed, Jul 29, 2020 at 10:51 AM Balaji K N <balaji.k.n at ericsson.com> wrote:
>> We are using OVS 2.12.0  with DPDK 18.11.2 release on Ubuntu 16.04 LTS. We do observe long initialization time for EAL init ( ~10 seconds for EAL memory initialization).
>>
>> Below is our hardware and memory configuration
>> Hardware                      : Dell Server
>> Sockets(NUMA)           : 2
>> CPUs                              : Each NUMA with 10 CPUs (Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz)
>> RAM                              : 128 GB
>> Hugepages reserverd : 2048 huge pages of 2MB and 100 Huge pages of 1G (reserved during boot time). We are using only 2MB huge pages for OVS process.
>> NIC                                  : Ethernet 10G 2P X520 ( ixgbe driver )
>>
>> Passing below EAL arguments with OVS process.
>> 2020-07-23T09:46:58.878Z|00014|dpdk|INFO|EAL ARGS: ovs-vswitchd -w 0000:04:00.0,support-multi-driver=1 -w 0000:42:00.0,support-multi-driver=1 -c 0x1 --huge-dir /mnt/huge_ovs_2M --socket-mem 2048,2048 --socket-limit 2048,2048.
>>
>> We debugged further in DPDK and found out that mmap called in eal_get_virtual_area is taking more time ( rte_eal_init->  rte_eal_memory_init -> rte_eal_memseg_init -> memseg_primary_init -> alloc_va_space -> eal_get_virtual_area ->mmap ). Long initialization time is noticeable difference compared to prior release. Any pointers would be helpful.
> 
> Did you try to start ovs with the --no-mlockall option?
> 
> 

That's most likely the culprit. I believe the recent mlock/MPROT changes are going to be backported to 18.11?

--
Thanks,
Anatoly