[OpenIndiana-discuss] RAM/CPU usage and ZFS send/receive

Nikola M minikola at gmail.com
Sun Apr 9 06:57:05 UTC 2017


On 04/ 8/17 03:44 PM, Harry Putnam wrote:
> I've noticed something like a several hundred GB send/recv with

Please see if you can make your subjects in ML messages On-topic, 
because I welcome general hardware talk, but we are actually more 
Openindiana/illumos Solaris-derived related then just hardware related.

ZFS usage of  RAM (ARC) can be controlled and if needed constrained by 
settings in (don't do it normally),
in /etc/system , with 'set zfs:zfs_arc_max =' and 'set zfs:zfs_arc_min 
=' in aether decimal or (0xnumber) hexadecimal form.
ZFS RAM usage can be bigger if you are using/have enabled per-dataset 
deduplication, where  (zfs dataset block sizes can be set 
per-dataset/pool), for every used deduplicated block one needs 
approximately 320 bytes of RAM.


> compression and the snap being transferred is lz4 on both ends.

zfs send always decompress the stream when sending and is recompressed 
on zfs receive.
Only if destination dataset is set to be compressed with compression 
property before receiving, then it recompress it on destination.
To speed up network transfer, one can , say use 'pigz' (parallel gzip) 
in between zfs stream send/receive, but is usually buffer also used 
before it, and also ssh to actually encrypt data stream over network.
(zfs send | mbuffer | pigz | ssh  /  ssh | pigz | mbuffer | zfs receive).

> The Host is pretty bogged down from that activity.
>
> It doesn't look so bad on the oi performance monitor (bld 151_9) with
> all 8 cores fluctuating from single digits up to about 32 % was the

Stop using 151a9, seriously. Only valid and supportable way of using 
Openindiana is at least using latest Openindiana 'snapshot' , to be able 
to report bugs ,if any, and/or update to state published /hipster 
rolling release IPS publisher. Or building one's own illumos if not 
wanting to use vanilla illumos coming through OI hipster.

Check if CPU/cores behave differently if EIST (frequency scaling on 
Intel CPU) is enabled or disabled
and wither you possibly have CPU consumed even when machine is idle or 
it is not.
I have machine where only way to stop using 30-40% of cpu time was to 
disable EIST, and I am not the only one.
But I was only able to see cpu use in GNOME/MATE System monitor in the 
panel.




More information about the openindiana-discuss mailing list