Re: [PATCH V1 net 2/2] net: ena: fix too long default tx interrupt moderation interval

["Machulsky, Zorik" <zorik@amazon.com>]

    
         > -----Original Message-----
        > From: David Miller <davem@davemloft.net>
        > Sent: Monday, November 4, 2019 9:19 PM
        > To: Kiyanovski, Arthur <akiyano@amazon.com>
        > Cc: netdev@vger.kernel.org; Woodhouse, David <dwmw@amazon.co.uk>;
        > Machulsky, Zorik <zorik@amazon.com>; Matushevsky, Alexander
        > <matua@amazon.com>; Bshara, Saeed <saeedb@amazon.com>; Wilson,
        > Matt <msw@amazon.com>; Liguori, Anthony <aliguori@amazon.com>;
        > Bshara, Nafea <nafea@amazon.com>; Tzalik, Guy <gtzalik@amazon.com>;
        > Belgazal, Netanel <netanel@amazon.com>; Saidi, Ali
        > <alisaidi@amazon.com>; Herrenschmidt, Benjamin <benh@amazon.com>;
        > Dagan, Noam <ndagan@amazon.com>; Agroskin, Shay
        > <shayagr@amazon.com>
        > Subject: Re: [PATCH V1 net 2/2] net: ena: fix too long default tx interrupt
        > moderation interval
        > 
        > From: <akiyano@amazon.com>
        > Date: Mon, 4 Nov 2019 13:58:48 +0200
        > 
        > > From: Arthur Kiyanovski <akiyano@amazon.com>
        > >
        > > Current default non-adaptive tx interrupt moderation interval is 196 us.
        > > This commit sets it to 0, which is much more sensible as a default value.
        > > It can be modified using ethtool -C.
        > >
        > > Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
        > 
        > I do not agree that turning TX interrupt moderation off completely is a more
        > sensible default value.
        > 
        > Maybe a much smaller value, but turning off the coalescing delay completely
        > is a bit much.
        
        David,
        Up until now, the ENA device did not support interrupt moderation, so effectively the default tx interrupt moderation interval was 0.
        You are probably right that 0 is not an optimal value.
        However until we research and find such an optimal value, in order to avoid a degradation in default performance, we want the default value in the new driver to be (effectively) the same as in the old driver.
        
     David,
    Just wanted to re-iterate what Arthur has mentioned. We clearly see BW and CPU utilization improvement with introduction of DIM on the Rx side and non-adaptive moderation on the Tx side in our driver. 
    We'd like to deliver this to our customers ASAP. We are usually very cautious with introduction of the new features, therefore we'd like to keep interrupt moderation disabled by default for now. We'd advise our customers awaiting for it to enable it using ethtool. 
    We are going to enable moderation by default after we accumulate more mileage with it and fine tune it further. That's the reason behind having Tx interval =0 for now (Rx moderation is disabled by default as well). Hope it makes sense.    

David,
I'm bumping this up to bring it to your attention.
Thanks.