[OpenIndiana-discuss] Arc and L2arc hit rate
Sebastian Gabler
sequoiamobil at gmx.net
Tue Mar 17 15:15:37 UTC 2015
Hi Geoff and Jason (and anybody else interested),
thanks for your comments.
I have been using arc_summary.pl. From what I understand the stats are
since boot time, which is around 700 days in my case, so the impact of
any tuning measures will not reflect there. Still thee are some findings:
- prefetch hit rate is rather low (8%). I am looking to speed up
sequential work loads. That's why I investigate in tuning ARC currently.
- indeed, zfs_arc_min ist at the default 12%, which is way below target.
Will increasing the value help with prefetch hit rates in my case, or
just set off the RAM ARC uses in general?
- are there any other measures that make sure that more prefetch will be
held in ARC without evicting MRU and MFU prematurely?
- can zfs_arc_min be set with mdb? (I won't reboot my machine for tuning
the ARC)
- is there a means to have prefetch going to the L2ARC directly,
bypassing ARC?
Lastly, can anybody explain the mechanism of ARC hit rates and L2ARC hit
rates? Do high ARC hit rates always mean low L2ARC hit rates, like I
assumed?
BR,
Sebastian
Am 12.03.2015 um 21:45 schrieb openindiana-discuss-request at openindiana.org:
> Hi Sebastian.
>
> Is it possible the working set of data is ~ 4.5GB in size?
>
> Also, you may want to use arc_summary as a tool to get a better
> understanding of your ARC usage
>
> http://www.cuddletech.com/arc_summary.pl
>
>
> Geoff
>
>
>
> ------------------------------
>
> Message: 4
> Date: Thu, 12 Mar 2015 10:14:24 -0700
> From: jason matthews<jason at broken.net>
> To:openindiana-discuss at openindiana.org
> Subject: Re: [OpenIndiana-discuss] Arc and L2arc hit rate
> Message-ID:<5501C970.9060108 at broken.net>
> Content-Type: text/plain; charset=windows-1252; format=flowed
>
>
> If you feel the need to micro manage the arc here are the parameters
> you'll like:
> zfs_arc_max
> zfs_arc_min
>
> you need/should to specify values in 64 bits. ie.
>
> * limit size of arc for hadoop systems
> set zfs:zfs_arc_max = 0x900000000
>
> or conversely
> set zfs:zfs_arc_min = 0x800000000
>
>
> understand that by nailing down the minimize size you are setting
> yourself up for memory starvation problems if you dont also tightly
> control memory usage. I dont use zfs_arc_min for this reason.
>
> here is the oracle page on the tuneable, for that is worth.
More information about the openindiana-discuss
mailing list