RE: [PATCH v2] net: netvsc: Update default VMBus channels

From: Michael Kelley
Date: Wed Aug 14 2024 - 19:57:23 EST


From: Erni Sri Satya Vennela <ernis@xxxxxxxxxxxxxxxxxxx> Sent: Wednesday, August 14, 2024 9:59 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 32 vCPUs, the channels
> default to VRSS_CHANNEL_DEFAULT. For greater than 32 vCPUs,
> set the channels to number of physical cores / 2 as a way
> to optimize CPU resource utilization and scale for high-end
> processors with many cores.
> Maximum number of channels are by default set to 64.

Where in the code is this enforced? It's not part of this patch. It
might be in rndis_set_subchannel(), where a value larger than
64 could be sent to the Hyper-V host, expecting that the Hyper-V
host will limit it to 64. But netvsc driver code is declaring an array
of size VRSS_CHANNEL_MAX, and there's nothing that guarantees
that Hyper-V will always limit the channel count to 64. But maybe
the netvsc driver enforces the limit of VRSS_CHANNEL_MAX in a
place that I didn't immediately see in a quick look at the code.

>
> Based on this change the subchannel creation would change as follows:
>
> -------------------------------------------------------------
> |No. of vCPU |dev_info->num_chn |subchannel created |
> -------------------------------------------------------------
> | 0-16 | 16 | vCPU |
> | >16 & <=32 | 16 | 16 |
> | >32 & <=128 | vCPU/2 | vCPU/2 |
> | >128 | vCPU/2 | 64 |
> -------------------------------------------------------------

The terminology here is slightly wrong. A VMBus device has one
primary channel plus zero or more subchannels. The chart
above is specifying the total number of channels (primary plus
subchannels), not the number of subchannels.

Michael

>
> Performance tests showed significant improvement in throughput:
> - 0.54% for 16 vCPUs
> - 0.83% for 32 vCPUs
> - 1.76% for 48 vCPUs
> - 10.35% for 64 vCPUs
> - 13.47% for 96 vCPUs
>
> Signed-off-by: Erni Sri Satya Vennela <ernis@xxxxxxxxxxxxxxxxxxx>
> ---
> 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 | 5 ++++-
> 2 files changed, 5 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..e32eb2997bf7 100644
> --- a/drivers/net/hyperv/netvsc_drv.c
> +++ b/drivers/net/hyperv/netvsc_drv.c
> @@ -27,6 +27,7 @@
> #include <linux/rtnetlink.h>
> #include <linux/netpoll.h>
> #include <linux/bpf.h>
> +#include <linux/cpumask.h>
>
> #include <net/arp.h>
> #include <net/route.h>
> @@ -987,7 +988,9 @@ struct netvsc_device_info *netvsc_devinfo_get(struct
> netvsc_device *nvdev)
> dev_info->bprog = prog;
> }
> } else {
> - dev_info->num_chn = VRSS_CHANNEL_DEFAULT;
> + int count = num_online_cpus();
> +
> + dev_info->num_chn = (count < 32) ? VRSS_CHANNEL_DEFAULT : DIV_ROUND_UP(count, 2);
> 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
>