ACK: [SRU][R:aws/Q:aws][PATCH 0/1] Set ENA_INTR_INITIAL_TX_INTERVAL_USECS to 64

Yufeng Gao yufeng.gao at canonical.com
Wed Mar 18 07:13:58 UTC 2026


On 18/3/26 05:32, Magali Lemes wrote:
> BugLink: https://bugs.launchpad.net/bugs/2144038
>
> [Impact]
> R:aws and Q:aws kernels set ENA_INTR_INITIAL_TX_INTERVAL_USECS to 0 in
> drivers/net/ethernet/amazon/ena/ena_com.h, disabling TX interrupt
> coalescing. This causes 10-25% Nginx performance regression compared to
> the Ubuntu 22.04 6.8 kernel, which correctly uses tx-usecs=64 matching
> the upstream default. This is likely to impact any network
> throughput-heavy application running on EC2.
>
> [Fix]
> Revert the SAUCE patch setting ENA_INTR_INITIAL_TX_INTERVAL_USECS to 0.
>
> [Test Plan]
> Launch an AWS EC2 instance that uses the ENA driver (e.g. m8g.2xlarge or
> m8i.2xlarge) running Ubuntu 24.04 to host the Nginx server. Launch a
> second instance to act as the load generator.
>
> On the server, boot a 6.17+ AWS kernel without the patch, verify
> tx-usecs is 0 (ethtool -c <interface-name> | grep tx-usecs), then
> generate load from the client using wrk (wrk -t 8 -c 1000 -d 60s
> http://<SERVER-IP>) and record the RPS. Again in the server, reboot into
> a patched 6.17+ AWS kernel, confirm tx-usecs is 64, and rerun the same
> workload. The patched kernel should show significantly higher RPS.
>
> [Where problems could occur]
> This change only affects the ENA driver, modifying the default TX
> interrupt coalescing value (tx-usecs) from 0 back to 64, which may
> introduce a small increase in packet transmission latency.
>
> Magali Lemes (1):
>    Revert "UBUNTU: SAUCE: net: ena: fix too long default tx interrupt
>      moderation interval"
>
>   drivers/net/ethernet/amazon/ena/ena_com.h | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
Acked-by: Yufeng Gao <yufeng.gao at canonical.com>




More information about the kernel-team mailing list