netdev - Re: [PATCH 10/11] nvmet-tcp: add NVMe over TCP target driver

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 19 Nov 2018 13:37:03 -0800
From:   Sagi Grimberg <sagi@...mberg.me>
To:     Max Gurtovoy <maxg@...lanox.com>,
        David Miller <davem@...emloft.net>, sagi@...htbitslabs.com
Cc:     linux-block@...r.kernel.org, netdev@...r.kernel.org,
        keith.busch@...el.com, hch@....de, linux-nvme@...ts.infradead.org
Subject: Re: [PATCH 10/11] nvmet-tcp: add NVMe over TCP target driver

>>> +static unsigned nvmet_tcp_recv_budget = 8;
>>> +module_param_named(recv_budget, nvmet_tcp_recv_budget, int, S_IRUGO 
>>> | S_IWUSR);
>>> +MODULE_PARM_DESC(recv_budget, "recvs budget");
>>> +
>>> +static unsigned nvmet_tcp_send_budget = 8;
>>> +module_param_named(send_budget, nvmet_tcp_send_budget, int, S_IRUGO 
>>> | S_IWUSR);
>>> +MODULE_PARM_DESC(send_budget, "sends budget");
>>> +
>>> +static unsigned nvmet_tcp_io_work_budget = 64;
>>> +module_param_named(io_work_budget, nvmet_tcp_io_work_budget, int, 
>>> S_IRUGO | S_IWUSR);
>>> +MODULE_PARM_DESC(io_work_budget, "io work budget");
>> I strongly suggest moving away from module parameters for this stuff.
> 
> agree here.
> 
> also, Sagi, can you explain about the performance trade-offs seen during 
> your development for these values ?
> 
> are they HCA/NIC dependent ?
> 
> should send/recv ratio be 1:1 ?
> 
> should total/send/recv ratio be 8:1:1 ?

These are not really HW dependent at all, its more about the tradeoff
between aggregation vs. fairness multiplexing. The budgets are designed
to control how much a specific workload (e.g. nvme queue) can hog the
cpu/wire when nvmet is servicing a large number of hosts.

There no constraints about the ratios of the budgets. Its advised though
that the io_work_budget would be able to catch at least a few
sends/recvs for reasonable aggregation.

I commented to Dave that I prefer not to expose them at this point given
that they are not trivial and would require an additional interface to
the driver (and its corresponding tool chain).