Looking at timestamped output of netdev CI reveals that most of the time in forwarding tests for custom route hashing is spent on a single case, namely the test which uses ping (mausezahn does not support flow labels).
On a non-debug kernel we spend 714 of 730 total test runtime (97%) on this test case. While having flow label support in a traffic gen tool / mausezahn would be best, we can significantly speed up the loop by putting ip vrf exec outside of the iteration.
In a test of 1000 pings using a normal loop takes 50 seconds to finish. While using:
ip vrf exec $vrf sh -c "$loop-body"
takes 12 seconds (1/4 of the time).
Some of the slowness is likely due to our inefficient virtualization setup, but even on my laptop running "ip link help" 16k times takes 25-30 seconds, so I think it's worth optimizing even for fastest setups.
Signed-off-by: Jakub Kicinski kuba@kernel.org --- CC: shuah@kernel.org CC: idosch@nvidia.com CC: linux-kselftest@vger.kernel.org --- .../selftests/net/forwarding/custom_multipath_hash.sh | 8 ++++---- .../selftests/net/forwarding/gre_custom_multipath_hash.sh | 8 ++++---- .../net/forwarding/ip6gre_custom_multipath_hash.sh | 8 ++++---- 3 files changed, 12 insertions(+), 12 deletions(-)
diff --git a/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh index 1783c10215e5..7d531f7091e6 100755 --- a/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh @@ -224,10 +224,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label. - for _ in $(seq 1 16384); do - ip vrf exec v$h1 \ - $PING6 2001:db8:4::2 -F 0 -c 1 -q >/dev/null 2>&1 - done + ip vrf exec v$h1 sh -c \ + "for _ in {1..16384}; do \ + $PING6 2001:db8:4::2 -F 0 -c 1 -q >/dev/null 2>&1; \ + done" }
send_src_udp6() diff --git a/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh index 9788bd0f6e8b..dda11a4a9450 100755 --- a/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh @@ -319,10 +319,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label. - for _ in $(seq 1 16384); do - ip vrf exec v$h1 \ - $PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1 - done + ip vrf exec v$h1 sh -c \ + "for _ in {1..16384}; do \ + $PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1; \ + done" }
send_src_udp6() diff --git a/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh index 2ab9eaaa5532..e28b4a079e52 100755 --- a/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh @@ -321,10 +321,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label. - for _ in $(seq 1 16384); do - ip vrf exec v$h1 \ - $PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1 - done + ip vrf exec v$h1 sh -c \ + "for _ in {1..16384}; do \ + $PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1; \ + done" }
send_src_udp6()
On Sat, Aug 17, 2024 at 01:36:59PM -0700, Jakub Kicinski wrote:
Looking at timestamped output of netdev CI reveals that most of the time in forwarding tests for custom route hashing is spent on a single case, namely the test which uses ping (mausezahn does not support flow labels).
On a non-debug kernel we spend 714 of 730 total test runtime (97%) on this test case. While having flow label support in a traffic gen tool / mausezahn would be best, we can significantly speed up the loop by putting ip vrf exec outside of the iteration.
In a test of 1000 pings using a normal loop takes 50 seconds to finish. While using:
ip vrf exec $vrf sh -c "$loop-body"
takes 12 seconds (1/4 of the time).
Some of the slowness is likely due to our inefficient virtualization setup, but even on my laptop running "ip link help" 16k times takes 25-30 seconds, so I think it's worth optimizing even for fastest setups.
Signed-off-by: Jakub Kicinski kuba@kernel.org
Reviewed-by: Ido Schimmel idosch@nvidia.com Tested-by: Ido Schimmel idosch@nvidia.com
Thanks!
On Sat, Aug 17, 2024 at 01:36:59PM -0700, Jakub Kicinski wrote:
Looking at timestamped output of netdev CI reveals that most of the time in forwarding tests for custom route hashing is spent on a single case, namely the test which uses ping (mausezahn does not support flow labels).
On a non-debug kernel we spend 714 of 730 total test runtime (97%) on this test case. While having flow label support in a traffic gen tool / mausezahn would be best, we can significantly speed up the loop by putting ip vrf exec outside of the iteration.
In a test of 1000 pings using a normal loop takes 50 seconds to finish. While using:
ip vrf exec $vrf sh -c "$loop-body"
takes 12 seconds (1/4 of the time).
Some of the slowness is likely due to our inefficient virtualization setup, but even on my laptop running "ip link help" 16k times takes 25-30 seconds, so I think it's worth optimizing even for fastest setups.
Signed-off-by: Jakub Kicinski kuba@kernel.org
CC: shuah@kernel.org CC: idosch@nvidia.com CC: linux-kselftest@vger.kernel.org
.../selftests/net/forwarding/custom_multipath_hash.sh | 8 ++++---- .../selftests/net/forwarding/gre_custom_multipath_hash.sh | 8 ++++---- .../net/forwarding/ip6gre_custom_multipath_hash.sh | 8 ++++---- 3 files changed, 12 insertions(+), 12 deletions(-)
diff --git a/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh index 1783c10215e5..7d531f7091e6 100755 --- a/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/custom_multipath_hash.sh @@ -224,10 +224,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label.
- for _ in $(seq 1 16384); do
ip vrf exec v$h1 \$PING6 2001:db8:4::2 -F 0 -c 1 -q >/dev/null 2>&1- done
- ip vrf exec v$h1 sh -c \
"for _ in {1..16384}; do \$PING6 2001:db8:4::2 -F 0 -c 1 -q >/dev/null 2>&1; \done"} send_src_udp6() diff --git a/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh index 9788bd0f6e8b..dda11a4a9450 100755 --- a/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/gre_custom_multipath_hash.sh @@ -319,10 +319,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label.
- for _ in $(seq 1 16384); do
ip vrf exec v$h1 \$PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1- done
- ip vrf exec v$h1 sh -c \
"for _ in {1..16384}; do \$PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1; \done"} send_src_udp6() diff --git a/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh b/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh index 2ab9eaaa5532..e28b4a079e52 100755 --- a/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh +++ b/tools/testing/selftests/net/forwarding/ip6gre_custom_multipath_hash.sh @@ -321,10 +321,10 @@ send_dst_ipv6() send_flowlabel() { # Generate 16384 echo requests, each with a random flow label.
- for _ in $(seq 1 16384); do
ip vrf exec v$h1 \$PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1- done
- ip vrf exec v$h1 sh -c \
"for _ in {1..16384}; do \$PING6 2001:db8:2::2 -F 0 -c 1 -q >/dev/null 2>&1; \done"} send_src_udp6() -- 2.46.0
Reviewed-by: Hangbin Liu liuhangbin@gmail.com
Hello:
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski kuba@kernel.org:
On Sat, 17 Aug 2024 13:36:59 -0700 you wrote:
Looking at timestamped output of netdev CI reveals that most of the time in forwarding tests for custom route hashing is spent on a single case, namely the test which uses ping (mausezahn does not support flow labels).
On a non-debug kernel we spend 714 of 730 total test runtime (97%) on this test case. While having flow label support in a traffic gen tool / mausezahn would be best, we can significantly speed up the loop by putting ip vrf exec outside of the iteration.
[...]
Here is the summary with links: - [net-next] selftests: net/forwarding: spawn sh inside vrf to speed up ping loop https://git.kernel.org/netdev/net-next/c/555e5531635a
You are awesome, thank you!
linux-kselftest-mirror@lists.linaro.org
-
Hangbin Liu -
Ido Schimmel -
Jakub Kicinski -
patchwork-bot+netdevbpf＠kernel.org