[Moosefs-users] auto-adjust to number of chunks per server ?

[Laurent W. <lw...@hy...>]

Hi,

About the packet too long thread, I'd like to talk a bit about possible
improvements before coding anything if needed.

MaxPacketSize could be dynamically adapted, as the +900 in « if
((uint32_t) (main_time())<=starttime+900) », and so is the 14400 in «
for (k=0 ; k<(NODEHASHSIZE/14400) && i<NODEHASHSIZE ; k++,i++) { ».

Border-line to this, MFSMAXFILES could also be dynamically allocated to
grow.

First, please note I'm unsure that basing auto-adjust on number of
chunks per server is the right way. If not, I guess you can trash
almost evething down there :)

I've looked quickly at the code, and it
seems quite easy to do. The only thing is to find some good enough basic
formulas so users don't or seldom meet these problems again.

Some numbers to try to find the way:

MaxPacketSize 50000000 works fine with 800000 chunks/server.
MaxPacketSize 50000000 does not work with 4700000 chunks/server.
MaxPacketSize 500000000 does work with 4700000 chunks/server.
50000000/800000=62.5 works
500000000/4700000~=106 works
50000000/4700000~=10.6 fails
I don't know the inner working of mfs enough, but it may work if
MaxPacketSize=number of chunks per server * 62.5, supposing
the evolution of MaxPacketSize only needs to be linear. If not, maybe
50000000+(number of chunks per server-800000)/~115 would do the trick ?
It would be great if fabien could try 62.5*4700000=293750000 for
MaxPacketSize. Or if moosefs main devs could bring their knowledge into
it.
BUT, there are several definitions to MaxPacketSize:
mfsmetalogger/masterconn.c  #define MaxPacketSize 1500000
mfsmaster/matocuserv.c      #define MaxPacketSize 1000000
mfsmaster/matomlserv.c      #define MaxPacketSize 1500000
mfsmaster/matocsserv.c      #define MaxPacketSize 50000000
mfschunkserver/masterconn.c #define MaxPacketSize 10000
mfschunkserver/csserv.c     #define MaxPacketSize 100000
Of course, here, only mfsmaster/matocsserv.c is concerned. I guess the
others are not as sensible about size or number of chunks.

About starttime+x
150 works fine with 800000 chunks/server.
900 is proposed with 4700000 chunks/server.
If linear is the way:
800k/150=5333.333…
4.7M/900=5222.222…
so x=number of chunks per server/5300 may do the trick ?
Or, if taking a minimum of 150, we'll get
x=150+(number of chunks per server-800k)/6000
with of course number of chunks-800k being≥0.

About NODEHASHSIZE/x
3600 works with 800k chunks per server.
14400 is proposed with 4.7M.
Linear:
800k/3600=222.222…
4.7M/14400=326.388…
x=number of chunks per server/222 ?
or with 3600 minimum:
x=3600+(number of chunks per server-800k)/361
with of course number of chunks-800k being≥0.

About MFSMAXFILES:
defaut is 5000 in code, Makefile pushes it to 10000.
I've found no information about tuning it.
Maybe we could simply use malloc for struct pollfd pdesc
[MFSMAXFILES]; and use realloc to MFSMAXFILES+something if « if
(setrlimit (RLIMIT_NOFILE,&rls)<0) { » is true ?

Thoughts ?
-- 
Laurent Wandrebeck
HYGEOS, Earth Observation Department / Observation de la Terre
Euratechnologies
165 Avenue de Bretagne
59000 Lille, France
tel: +33 3 20 08 24 98
http://www.hygeos.com
GPG fingerprint/Empreinte GPG: F5CA 37A4 6D03 A90C 7A1D  2A62 54E6 EF2C
D17C F64C