From: Erni Sri Satya Vennela <ernis@xxxxxxxxxxxxxxxxxxx> Sent: Thursday, August 22, 2024 8:06 AM
>
> Change VMBus channels macro (VRSS_CHANNEL_DEFAULT) in
> Linux netvsc from 8 to 16 to align with Azure Windows VM
> and improve networking throughput.
>
> For VMs having less than 16 vCPUS, the channels depend
> on number of vCPUs. Between 16 to 64 vCPUs, the channels
> default to VRSS_CHANNEL_DEFAULT. For greater than 64 vCPUs,
> set the channels to number of physical cores / 2 returned by
> netif_get_num_default_rss_queues() as a way to optimize CPU
> resource utilization and scale for high-end processors with
> many cores. Due to hyper-threading, the number of
> physical cores = vCPUs/2.
But note that a given physical processor may or may not support
hyper-threading. For example, the physical processor used for
ARM64 VMs in Azure does not have hyper-threading. And even
if the physical processor supports hyper-threading, the VM might
not see hyper-threading as enabled. Many Azure GPU-based VM
sizes see only full cores, with no hyper-threading. It's also possible
to boot Linux with hyper-threading disabled even if the VM sees
hyper-threaded cores (the "nosmt" or "smt=1" kernel boot option).
Your code below probably isn't affected when hyper-threading
isn't present. But in the interest of accuracy, the discussion here
in the commit message should qualify the use of "vCPU/4" as
the number of channels. It might be "vCPU/2" when
hyper-threading isn't present or is disabled, and for vCPU
counts between 16 and 64, you'll get more than 16 channels.
> Maximum number of channels are by default set to 64.
>
> Based on this change the channel creation would change as follows:
>
> -------------------------------------------------------------
> | No. of vCPU | dev_info->num_chn | channels created |
> -------------------------------------------------------------
> | 0-16 | 16 | vCPU |
Nit: Presumably we won't ever have 0 vCPUs. :-)
> | >16 & <=64 | 16 | 16 |
> | >64 & <=256 | vCPU/4 | vCPU/4 |
> | >256 | vCPU/4 | 64 |
> -------------------------------------------------------------
>
> Performance tests showed significant improvement in throughput:
> - 0.54% for 16 vCPUs
> - 0.83% for 32 vCPUs
> - 0.86% for 48 vCPUs
> - 9.72% for 64 vCPUs
> - 13.57% for 96 vCPUs
>
> Signed-off-by: Erni Sri Satya Vennela <ernis@xxxxxxxxxxxxxxxxxxx>
> Reviewed-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> Reviewed-by: Shradha Gupta <shradhagupta@xxxxxxxxxxxxxxxxxxx>
> ---
> Changes in v3:
> * Use netif_get_num_default_rss_queues() to set channels
> * Change terminology for channels in commit message
> ---
> Changes in v2:
> * Set dev_info->num_chn based on vCPU count.
> ---
> drivers/net/hyperv/hyperv_net.h | 2 +-
> drivers/net/hyperv/netvsc_drv.c | 3 ++-
> 2 files changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/hyperv/hyperv_net.h b/drivers/net/hyperv/hyperv_net.h
> index 810977952f95..e690b95b1bbb 100644
> --- a/drivers/net/hyperv/hyperv_net.h
> +++ b/drivers/net/hyperv/hyperv_net.h
> @@ -882,7 +882,7 @@ struct nvsp_message {
>
> #define VRSS_SEND_TAB_SIZE 16 /* must be power of 2 */
> #define VRSS_CHANNEL_MAX 64
> -#define VRSS_CHANNEL_DEFAULT 8
> +#define VRSS_CHANNEL_DEFAULT 16
>
> #define RNDIS_MAX_PKT_DEFAULT 8
> #define RNDIS_PKT_ALIGN_DEFAULT 8
> diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
> index 44142245343d..a6482afe4217 100644
> --- a/drivers/net/hyperv/netvsc_drv.c
> +++ b/drivers/net/hyperv/netvsc_drv.c
> @@ -987,7 +987,8 @@ struct netvsc_device_info *netvsc_devinfo_get(struct netvsc_device *nvdev)
> dev_info->bprog = prog;
> }
> } else {
> - dev_info->num_chn = VRSS_CHANNEL_DEFAULT;
> + dev_info->num_chn = max(VRSS_CHANNEL_DEFAULT,
> + netif_get_num_default_rss_queues());
> dev_info->send_sections = NETVSC_DEFAULT_TX;
> dev_info->send_section_size = NETVSC_SEND_SECTION_SIZE;
> dev_info->recv_sections = NETVSC_DEFAULT_RX;
> --
> 2.34.1
>
