net_dim.h lib exposes an implementation of the DIM algorithm for dynamically-tuned interrupt moderation for networking interfaces. We need the same behavior for any block CQ. The main motivation is two benefit from maximized completion rate and reduced interrupt overhead that DIM may provide. Current DIM implementation prioritizes reducing interrupt overhead over latency. Also, in order to reduce DIM's own overhead, the algorithm might take take some time to identify it needs to change profiles. For these reasons we got to the understanding that a slightly modified algorithm is needed. Early tests with current implementation show it doesn't react fast and sharply enough in order to satisfy the block CQ needs. I would like to suggest an implementation for block DIM. The idea is to expose the new functionality without the risk of breaking Net DIM behavior for netdev. Below are main similarities and differences between the two implementations and general guidelines for the suggested solution. Performance tests over ConnectX-5 100GbE NIC show a 200% improvement on tail latency when switching from high load traffic to low load traffic. Common logic, main DIM procedure: - Calculate current stats from a given sample - Compare current stats vs. previous iteration stats - Make a decision -> choose a new profile Differences: - Different parameters for moving between profiles - Different moderation values and number of profiles - Different sampled data Suggested solution: - Common logic will be declared in include/linux/dim.h and implemented in lib/dim/dim.c - Net DIM (existing) logic will be declared in include/linux/net_dim.h and implemented in lib/dim/net_dim.c, which will use the common logic from dim.h - Block DIM logic will be declared in /include/linux/block_dim.h and implemented in lib/dim/blk_dim.c. This new implementation will expose modified versions of profiles, dim_step() and dim_decision() Pros for this solution are: - Zero impact on existing net_dim implementation and usage - Relatively more code reuse (compared to two separate solutions) - Readiness for future implementations Tal Gilboa (6): linux/dim: Move logic to dim.h linux/dim: Remove "net" prefix from internal DIM members linux/dim: Rename externally exposed macros linux/dim: Rename net_dim_sample() to net_dim_create_sample() linux/dim: Rename externally used net_dim members linux/dim: Move implementation to .c files Yamin Friedman (3): linux/dim: Add completions count to dim_sample linux/dim: Implement blk_dim.h drivers/infiniband: Use blk_dim in infiniband driver MAINTAINERS | 3 + drivers/infiniband/core/cq.c | 75 +++- drivers/infiniband/hw/mlx4/qp.c | 2 +- drivers/infiniband/hw/mlx5/qp.c | 2 +- drivers/net/ethernet/broadcom/bcmsysport.c | 20 +- drivers/net/ethernet/broadcom/bcmsysport.h | 2 +- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 13 +- drivers/net/ethernet/broadcom/bnxt/bnxt.h | 2 +- .../net/ethernet/broadcom/bnxt/bnxt_debugfs.c | 4 +- drivers/net/ethernet/broadcom/bnxt/bnxt_dim.c | 7 +- .../net/ethernet/broadcom/genet/bcmgenet.c | 18 +- .../net/ethernet/broadcom/genet/bcmgenet.h | 2 +- drivers/net/ethernet/mellanox/mlx5/core/en.h | 8 +- .../net/ethernet/mellanox/mlx5/core/en_dim.c | 12 +- .../ethernet/mellanox/mlx5/core/en_ethtool.c | 4 +- .../net/ethernet/mellanox/mlx5/core/en_main.c | 22 +- .../net/ethernet/mellanox/mlx5/core/en_txrx.c | 10 +- include/linux/blk_dim.h | 56 +++ include/linux/dim.h | 126 +++++++ include/linux/irq_poll.h | 7 + include/linux/net_dim.h | 338 +----------------- include/rdma/ib_verbs.h | 11 +- lib/Kconfig | 7 + lib/Makefile | 1 + lib/dim/Makefile | 14 + lib/dim/blk_dim.c | 114 ++++++ lib/dim/dim.c | 92 +++++ lib/dim/net_dim.c | 193 ++++++++++ lib/irq_poll.c | 13 +- 29 files changed, 778 insertions(+), 400 deletions(-) create mode 100644 include/linux/blk_dim.h create mode 100644 include/linux/dim.h create mode 100644 lib/dim/Makefile create mode 100644 lib/dim/blk_dim.c create mode 100644 lib/dim/dim.c create mode 100644 lib/dim/net_dim.c -- 2.19.1
