Gandi.net Groups

In a recent blog post, Gandi aknowledged that the system has memory
issues:

"In addition, if your server only has one share, we have corrected a
minor problem of excessive memory usage. To fix this, you only need to
reboot your server."

In my experience, this is not the case. I started tracking this when the
below appeared on my 1 share server with Gandi AI. The server was at
that point only a web server, no mail, and was basically dormant with
regards to traffic:

#15.06.2008
#              total       used       free     shared    buffers    
cached
# Mem:           356        211        144          0         67        
53
# -/+ buffers/cache:         90        265
# Swap:          619          0        619

I mysteriously gained some RAM when the 1-share limit was almost eaten
up. I have since upgraded this server 2 shares and put a site on there
that generates a modest amount of traffic. Still no mail, only a web
server. The results below show that the server incrementally eats around
a MB or so of RAM per hour:

# 25.06.2008 at 09.30
#              total       used       free     shared    buffers    
cached
# Mem:           512        225        286          0         12       
134
# -/+ buffers/cache:         78        433
# Swap:          619          0        619

# 25.06.2008 at 10.19
#              total       used       free     shared    buffers    
cached
# Mem:           512        226        285          0         12       
134
# -/+ buffers/cache:         78        433
# Swap:          619          0        619

# 25.06.2008 at 13.27
#              total       used       free     shared    buffers    
cached
# Mem:           512        232        280          0         16       
136
# -/+ buffers/cache:         79        432
# Swap:          619          0        619

# 26.06.2008 at 15.47
#              total       used       free     shared    buffers    
cached
# Mem:           512        316        196          0         73       
154
# -/+ buffers/cache:         88        423
# Swap:          619          0        619

# 26.06.2008 at 17.30
#              total       used       free     shared    buffers    
cached
# Mem:           512        311        200          0         74       
154
# -/+ buffers/cache:         82        429
# Swap:          619          0        619

# 27.06.2008 at 09.32
#              total       used       free     shared    buffers    
cached
# Mem:           512        327        184          0         80       
158
# -/+ buffers/cache:         88        423
# Swap:          619          0        619

# 29.06.2008 at 12.06
#              total       used       free     shared    buffers    
cached
# Mem:           512        447         64          0         83       
282
# -/+ buffers/cache:         81        430
# Swap:          619          0        619

Running "top" at all these instances show almost identical results with
very low resource usage. I would be curious if other users of the
platform, especially people with more shares, are seeing similar results
and if Gandi perhaps has some input on the situation. I realize of
course that memory usage usually increases over time and that a reboot
resets the counter, but this appears to be more of a flood than a leak.

On Jun, 29 2008 12:25 CEST, mc wrote:

#15.06.2008
# -/+ buffers/cache:         90
# 25.06.2008 at 09.30
# -/+ buffers/cache:         78
# 25.06.2008 at 10.19
# -/+ buffers/cache:         78
# 25.06.2008 at 13.27
# -/+ buffers/cache:         79
# 26.06.2008 at 15.47
# -/+ buffers/cache:         88
# 26.06.2008 at 17.30
# -/+ buffers/cache:         82
# 27.06.2008 at 09.32  
# -/+ buffers/cache:         88
# 29.06.2008 at 12.06
# -/+ buffers/cache:         81

Voilà!

OK, I get that usage has been constant and that there is plenty spare.
The concern was primarily what happens when nearly all the physical
memory is reserved by the system. Even on a 512 MB server, it seems. I
suspect that Gandi offered the extra 100 MB of physical memory for the 1
share to resolve an issue with the system not releasing it properly. 

As the numbers state, the server is obviously lightning fast at the
moment. So no complaints about performance, simply trying to understand
the memory issues and avoid any problems that may arise.

There is absolutely no need to free buffers with so much totaly unused
memory around. 

99.9% memory usage is the most perfect runtime state, as much as
possible is hit through the buffer. Why not use the (very fast) memory
as much as possible?

Even 99.9% swap usage would not hurt if pages are swapped in/out very
infrequently.

Just watch how often swap and unbuffered-disk are hit, which can (but
might not) both be lowered with more memory. Disk is awfully very very
very very much slower than memory, so you want to use it as infrequent
as possible. Since memory is rather expecive, decide wisely... a server
just pumping out files from disk to network is probably also limited by
its network-link (very true for the gandi-case), so more memory will
never speed it up. Find your personal bottleneck and increase power
there first.


(not universially true, but for the quick and gerneral explanaition,
this should give you the right hints)

Thank you.