Re: CURRENT slow and shaky network stability

Am Sat, 2 Apr 2016 11:39:10 +0200
"O. Hartmann" <ohartman_at_zedat.fu-berlin.de> schrieb:

> Am Sat, 2 Apr 2016 10:55:03 +0200
> "O. Hartmann" <ohartman_at_zedat.fu-berlin.de> schrieb:
> 
> > Am Sat, 02 Apr 2016 01:07:55 -0700
> > Cy Schubert <Cy.Schubert_at_komquats.com> schrieb:
> >   
> > > In message <56F6C6B0.6010103_at_protected-networks.net>, Michael Butler writes:    
> > > > -current is not great for interactive use at all. The strategy of
> > > > pre-emptively dropping idle processes to swap is hurting .. big time.      
> > > 
> > > FreeBSD doesn't "preemptively" or arbitrarily push pages out to disk. LRU 
> > > doesn't do this.
> > >     
> > > > 
> > > > Compare inactive memory to swap in this example ..
> > > > 
> > > > 110 processes: 1 running, 108 sleeping, 1 zombie
> > > > CPU:  1.2% user,  0.0% nice,  4.3% system,  0.0% interrupt, 94.5% idle
> > > > Mem: 474M Active, 1609M Inact, 764M Wired, 281M Buf, 119M Free
> > > > Swap: 4096M Total, 917M Used, 3178M Free, 22% Inuse      
> > > 
> > > To analyze this you need to capture vmstat output. You'll see the free pool 
> > > dip below a threshold and pages go out to disk in response. If you have 
> > > daemons with small working sets, pages that are not part of the working 
> > > sets for daemons or applications will eventually be paged out. This is not 
> > > a bad thing. In your example above, the 281 MB of UFS buffers are more 
> > > active than the 917 MB paged out. If it's paged out and never used again, 
> > > then it doesn't hurt. However the 281 MB of buffers saves you I/O. The 
> > > inactive pages are part of your free pool that were active at one time but 
> > > now are not. They may be reclaimed and if they are, you've just saved more 
> > > I/O.
> > > 
> > > Top is a poor tool to analyze memory use. Vmstat is the better tool to help 
> > > understand memory use. Inactive memory isn't a bad thing per se. Monitor 
> > > page outs, scan rate and page reclaims.
> > > 
> > >     
> > 
> > I give up! Tried to check via ssh/vmstat what is going on. Last lines before broken
> > pipe:
> > 
> > [...]
> > procs  memory       page                    disks     faults         cpu
> > r b w  avm   fre   flt  re  pi  po    fr   sr ad0 ad1   in    sy    cs us sy id
> > 22 0 22 5.8G  1.0G 46319   0   0   0 55721 1297   0   4  219 23907  5400 95  5  0
> > 22 0 22 5.4G  1.3G 51733   0   0   0 72436 1162   0   0  108 40869  3459 93  7  0
> > 15 0 22  12G  1.2G 54400   0  27   0 52188 1160   0  42  148 52192  4366 91  9  0
> > 14 0 22  12G  1.0G 44954   0  37   0 37550 1179   0  39  141 86209  4368 88 12  0
> > 26 0 22  12G  1.1G 60258   0  81   0 69459 1119   0  27  123 779569 704359 87 13  0
> > 29 3 22  13G  774M 50576   0  68   0 32204 1304   0   2  102 507337 484861 93  7  0
> > 27 0 22  13G  937M 47477   0  48   0 59458 1264   3   2  112 68131 44407 95  5  0
> > 36 0 22  13G  829M 83164   0   2   0 82575 1225   1   0  126 99366 38060 89 11  0
> > 35 0 22 6.2G  1.1G 98803   0  13   0 121375 1217   2   8  112 99371  4999 85 15  0
> > 34 0 22  13G  723M 54436   0  20   0 36952 1276   0  17  153 29142  4431 95  5  0
> > Fssh_packet_write_wait: Connection to 192.168.0.1 port 22: Broken pipe
> > 
> > 
> > This makes this crap system completely unusable. The server (FreeBSD 11.0-CURRENT #20
> > r297503: Sat Apr  2 09:02:41 CEST 2016 amd64) in question did poudriere bulk job. I
> > can not even determine what terminal goes down first - another one, much more time
> > idle than the one shwoing the "vmstat 5" output, is still alive! 
> > 
> > i consider this a serious bug and it is no benefit what happened since this "fancy"
> > update. :-(  
> 
> By the way - it might be of interest and some hint.
> 
> One of my boxes is acting as server and gateway. It utilises NAT, IPFW, when it is under
> high load, as it was today, sometimes passing the network flow from ISP into the network
> for clients is extremely slow. I do not consider this the reason for collapsing ssh
> sessions, since this incident happens also under no-load, but in the overall-view onto
> the problem, this could be a hint - I hope. 

I just checked on one box, that "broke pipe" very quickly after I started poudriere,
while it did well a couple of hours before until the pipe broke. It seems it's load
dependend when the ssh session gets wrecked, but more important, after the long-haul
poudriere run, I rebooted the box and tried again with the mentioned broken pipe after a
couple of minutes after poudriere ran. Then I left the box for several hours and logged
in again and checked the swap. Although there was for hours no load or other pressure,
there were 31% of of swap used - still (box has 16 GB of RAM and is propelled by a XEON
E3-1245 V2).