Constraining CPU usage
Nash Nipples
trashy_bumper at yahoo.com
Sat Dec 3 07:18:40 PST 2005
Arne Woerner <arne_woerner at yahoo.com> wrote: --- Shane Ambler wrote:
> On 3/12/2005 3:14, "Arne Woerner"
> wrote:
> > Why should somebody want to keep CPU usage of a
> > process below a certain value (e. g. 20%)?
>
> Well if your machine isn't smp enabled (specifically
> pre-HTT) then one process can slow the server down
> to a point where your web/databse server can't get
> processor time to give a response - effectively making
> it non existent for the duration.
>
Hmm...
Isn't that problem solved already by nice'ing a process?
I think that problem is completely solved by idprio'ing a
process...
Because:
When the NIC gets a packet for a database server, the kernel
interrupts the executing of a potential idprio'ed process and
starts the DB process... Or am I wrong?
-Arne
Looks pretty simple and clear. Thank you Arne.
In respect to what Mike said i must admit: "its all about high utilization numbers fear"
If the machine was smp enabled i think the number would be around 50% or so.. thank you Shane.
Reasoning the topic i believe i should do a little research on:
Understanding units and measurements used in FreeBSD performance representation.
Calculating a machine's available resources and estimating its possible low/average/peak load when performing user/system tasks during a specific amount of time (24 hours). E.g: 1 user will read $File of $Size Mb from $Storage_media using $Transfer_media. The operation will last $time and consume $CPU CPU time, $Storage_media_capacity $Transfer_media_bandwith, $RAM, $Other shared resources.
So eventually, one can build an operations chart/table, estimate timings and balance a target machine's load among the processes that are a subject to order/priority.
Gather and analyze collected in a real-time situation statistics.
I'm planning to google that for a while.
Thank you everyone once again.
nash
---------------------------------
Yahoo! DSL Something to write home about. Just $16.99/mo. or less
More information about the freebsd-performance
mailing list