[v5,bpf-next,0/5] bpf_prog_pack followup

Message

Song Liu June 24, 2022, 9:57 p.m. UTC

This set is the second half of v4 [1].

Changes v4 => v5:
1. Rebase and resolve conflicts due to module.c split.
2. Update experiment results (below).

For our web service production benchmark, bpf_prog_pack on 4kB pages
gives 0.5% to 0.7% more throughput than not using bpf_prog_pack.
bpf_prog_pack on 2MB pages 0.6% to 0.9% more throughput than not using
bpf_prog_pack. Note that 0.5% is a huge improvement for our fleet. I
believe this is also significant for other companies with many thousand
servers.

Update: Further experiments (suggested by Rick Edgecombe) showed that most
of benefit on the web service benchmark came from less direct map
fragmentation. The experiment is as follows:

Side A: 2MB bpf prog pack on a single 2MB page;
Side B: 2MB bpf prog pack on 512x 4kB pages;

The system only uses about 200kB for BPF programs, but 2MB is allocated
for bpf_prog_pack (for both A and B). Therefore, direct map fragmentation
caused by BPF programs is elminated, and we are only measuring the
performance difference of 1x 2MB page vs. ~50 4kB pages (we only use
about 50 out of the 512 pages). For these two sides, the difference in
system throughput is within the noise. I also measured iTLB-load-misses
caused by bpf programs, which is ~300/s for case A, and ~1600/s for case B.
The overall iTLB-load-misses is about 1.5M/s on these hosts. Therefore,
we can clearly see 2MB page reduces iTLB misses, but the difference is not
enough to have visible impact on system throughput.

Of course, the impact of iTLB miss will be more significant for systems
with more BPF programs loaded.

[1] https://lore.kernel.org/bpf/[e-mail]/

Song Liu (5):
  module: introduce module_alloc_huge
  bpf: use module_alloc_huge for bpf_prog_pack
  vmalloc: WARN for set_vm_flush_reset_perms() on huge pages
  vmalloc: introduce huge_vmalloc_supported
  bpf: simplify select_bpf_prog_pack_size

 arch/x86/kernel/module.c     | 21 +++++++++++++++++++++
 include/linux/moduleloader.h |  5 +++++
 include/linux/vmalloc.h      |  7 +++++++
 kernel/bpf/core.c            | 25 ++++++++++---------------
 kernel/module/main.c         |  8 ++++++++
 mm/vmalloc.c                 |  5 +++++
 6 files changed, 56 insertions(+), 15 deletions(-)

--
2.30.2

Comments

Song Liu June 24, 2022, 10 p.m. UTC | #1

oops, wrong address for x86@. 

CC [e-mail]

> On Jun 24, 2022, at 2:57 PM, Song Liu <[e-mail]> wrote:
> 
> This set is the second half of v4 [1].
> 
> Changes v4 => v5:
> 1. Rebase and resolve conflicts due to module.c split.
> 2. Update experiment results (below).
> 
> For our web service production benchmark, bpf_prog_pack on 4kB pages
> gives 0.5% to 0.7% more throughput than not using bpf_prog_pack.
> bpf_prog_pack on 2MB pages 0.6% to 0.9% more throughput than not using
> bpf_prog_pack. Note that 0.5% is a huge improvement for our fleet. I
> believe this is also significant for other companies with many thousand
> servers.
> 
> Update: Further experiments (suggested by Rick Edgecombe) showed that most
> of benefit on the web service benchmark came from less direct map
> fragmentation. The experiment is as follows:
> 
> Side A: 2MB bpf prog pack on a single 2MB page;
> Side B: 2MB bpf prog pack on 512x 4kB pages;
> 
> The system only uses about 200kB for BPF programs, but 2MB is allocated
> for bpf_prog_pack (for both A and B). Therefore, direct map fragmentation
> caused by BPF programs is elminated, and we are only measuring the
> performance difference of 1x 2MB page vs. ~50 4kB pages (we only use
> about 50 out of the 512 pages). For these two sides, the difference in
> system throughput is within the noise. I also measured iTLB-load-misses
> caused by bpf programs, which is ~300/s for case A, and ~1600/s for case B.
> The overall iTLB-load-misses is about 1.5M/s on these hosts. Therefore,
> we can clearly see 2MB page reduces iTLB misses, but the difference is not
> enough to have visible impact on system throughput.
> 
> Of course, the impact of iTLB miss will be more significant for systems
> with more BPF programs loaded.
> 
> [1] https://lore.kernel.org/bpf/[e-mail]/
> 
> Song Liu (5):
>  module: introduce module_alloc_huge
>  bpf: use module_alloc_huge for bpf_prog_pack
>  vmalloc: WARN for set_vm_flush_reset_perms() on huge pages
>  vmalloc: introduce huge_vmalloc_supported
>  bpf: simplify select_bpf_prog_pack_size
> 
> arch/x86/kernel/module.c     | 21 +++++++++++++++++++++
> include/linux/moduleloader.h |  5 +++++
> include/linux/vmalloc.h      |  7 +++++++
> kernel/bpf/core.c            | 25 ++++++++++---------------
> kernel/module/main.c         |  8 ++++++++
> mm/vmalloc.c                 |  5 +++++
> 6 files changed, 56 insertions(+), 15 deletions(-)
> 
> --
> 2.30.2