[dpdk-users] Slow DPDK startup with many 1G hugepages

[Tan, Jianfeng]

> -----Original Message-----
> From: users [mailto:users-bounces at dpdk.org] On Behalf Of Imre Pinter
> Sent: Thursday, June 1, 2017 3:55 PM
> To: users at dpdk.org
> Cc: Gabor Halász; Péter Suskovics
> Subject: [dpdk-users] Slow DPDK startup with many 1G hugepages
> 
> Hi,
> 
> We experience slow startup time in DPDK-OVS, when backing memory with
> 1G hugepages instead of 2M hugepages.
> Currently we're mapping 2M hugepages as memory backend for DPDK OVS.
> In the future we would like to allocate this memory from the 1G hugepage
> pool. Currently in our deployments we have significant amount of 1G
> hugepages allocated (min. 54G) for VMs and only 2G memory on 2M
> hugepages.
> 
> Typical setup for 2M hugepages:
>                 GRUB:
> hugepagesz=2M hugepages=1024 hugepagesz=1G hugepages=54
> default_hugepagesz=1G
> 
> $ grep hugetlbfs /proc/mounts
> nodev /mnt/huge_ovs_2M hugetlbfs rw,relatime,pagesize=2M 0 0
> nodev /mnt/huge_qemu_1G hugetlbfs rw,relatime,pagesize=1G 0 0
> 
> Typical setup for 1GB hugepages:
> GRUB:
> hugepagesz=1G hugepages=56 default_hugepagesz=1G
> 
> $ grep hugetlbfs /proc/mounts
> nodev /mnt/huge_qemu_1G hugetlbfs rw,relatime,pagesize=1G 0 0
> 
> DPDK OVS startup times based on the ovs-vswitchd.log logs:
> 
>   *   2M (2G memory allocated) - startup time ~3 sec:
> 
> 2017-05-03T08:13:50.177Z|00009|dpdk|INFO|EAL ARGS: ovs-vswitchd -c 0x1
> --huge-dir /mnt/huge_ovs_2M --socket-mem 1024,1024
> 
> 2017-05-03T08:13:50.708Z|00010|ofproto_dpif|INFO|netdev at ovs-netdev:
> Datapath supports recirculation
> 
>   *   1G (56G memory allocated) - startup time ~13 sec:
> 2017-05-03T08:09:22.114Z|00009|dpdk|INFO|EAL ARGS: ovs-vswitchd -c 0x1
> --huge-dir /mnt/huge_qemu_1G --socket-mem 1024,1024
> 2017-05-03T08:09:32.706Z|00010|ofproto_dpif|INFO|netdev at ovs-netdev:
> Datapath supports recirculation
> I used DPDK 16.11 for OVS and testpmd and tested on Ubuntu 14.04 with
> kernel 3.13.0-117-generic and 4.4.0-78-generic.


You can shorten the time by this:

(1) Mount 1 GB hugepages into two directories.
nodev /mnt/huge_ovs_1G hugetlbfs rw,relatime,pagesize=1G,size=<how much you want to use in OVS> 0 0
nodev /mnt/huge_qemu_1G hugetlbfs rw,relatime,pagesize=1G 0 0

(2) Force to use memory  interleave policy 
$ numactl --interleave=all ovs-vswitchd ...

Note: keep the huge-dir and socket-mem option, "--huge-dir /mnt/huge_ovs_1G --socket-mem 1024,1024".

> 
> We had a discussion with Mark Gray (from Intel), and he come up with the
> following items:
> 
> ·         The ~10 sec time difference is there with testpmd as well
> 
> ·         They believe it is a kernel overhead (mmap is slow, perhaps it is zeroing
> pages). The following code from eal_memory.c does the above mentioned
> printout in EAL startup:

Yes, correct.

> 469    /* map the segment, and populate page tables,
> 470     * the kernel fills this segment with zeros */
> 468    uint64_t start = rte_rdtsc();
> 471    virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
> 472                    MAP_SHARED | MAP_POPULATE, fd, 0);
> 473    if (virtaddr == MAP_FAILED) {
> 474            RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
> 475                            strerror(errno));
> 476            close(fd);
> 477            return i;
> 478    }
> 479
> 480    if (orig) {
> 481            hugepg_tbl[i].orig_va = virtaddr;
> 482            printf("Original mapping of page %u took: %"PRIu64"
> ticks, %"PRIu64" ms\n     ",
> 483                    i, rte_rdtsc() - start,
> 484                    (rte_rdtsc() - start) * 1000 /
> 485                    rte_get_timer_hz());
> 486    }
> 
> 
> A solution could be to mount 1G hugepages to 2 separate directory: 2G for
> OVS and the remaining for the VMs, but the NUMA location for these
> hugepages is non-deterministic. Since mount cannot handle NUMA related
> parameters during mounting hugetlbfs, and fstab forks the mounts during
> boot.

Oh, similar idea :-)

> 
> Do you have a solution on how to use 1G hugepages for VMs and have
> reasonable DPDK EAL startup time?

No, we still don't have such options.

Thanks,
Jianfeng

> 
> Thanks,
> Imre