Two Itanium systems: SSH works one place, fails another

[Jack Baruth_1]

We have two near-identical Itanium boxes. On one OpenSSH 3.9 works like a charm; on the other, connections are continually dropped and the daemon dies. Moved to 4.2, connections still drop but the daemons live. Diffed the kernel on both boxes; working box to left, broken box to right. Any ideas?

< maxfiles_lim 4096 4096 Immed
---
> maxfiles_lim 2048 2048 Immed
62c62
< maxuprc 3360 ((nproc*8)/10) Immed
---
> maxuprc 1654 ((nproc*8)/10) Immed
64c64
< msgmap 4202 (msgtql+2)
---
> msgmap 2070 (msgtql+2)
67,68c67,68
< msgmni 4200 (nproc)
< msgseg 16800 (msgtql*4)
---
> msgmni 2068 (nproc)
> msgseg 8272 (msgtql*4)
70c70
< msgtql 4200 (nproc)
---
> msgtql 2068 (nproc)
74,78c74,78
< nfile 65536 65536 Imm (auto disabled)
< nflocks 4200 (nproc) Imm (auto disabled)
< ninode 35648 (8*nproc+2048)
< nkthread 8416 8416 Immed
< nproc 4200 4200 Immed
---
> nfile 30000 30000 Imm (auto disabled)
> nflocks 2068 (nproc) Imm (auto disabled)
> ninode 18592 (8*nproc+2048)
> nkthread 6000 6000 Immed
> nproc 2068 2068 Immed
91c91
< physical_io_buffers 1280 Default Auto
---
> physical_io_buffers 640 Default Auto
97c97
< scroll_lines 100 Default Immed
---
> scroll_lines 100 Default
104c104
< semmnu 4196 (nproc-4)
---
> semmnu 2064 (nproc-4)

[Mel Burslan]

I am not an expert on kernel tuning but your problematic server's kernel parameters looks lie a little too low. If the two servers identical hardware wise, why not bring them up to the same values ?

[Jack Baruth_1]

I agree with you and we are bringing them into line.. but when I look at what the parameters actually are, I don't see how any of them could cause trouble on what is basically an idle server.

[Mel Burslan]

if you have glance installed, did you try peeking at the cpu/memory usage, when firing up the sshd and in time how it perfors as people come in and out using ssh ? As I said, I can not relate any of the parameters to any specific activity on a server which is otherwise idle, but in some cases, smaller than expected values, cause some unforeseen ill effects. I haved lived thru quite few of those, not necessarily ssh related though. If you have a full glance/perfview license, you can analyze the historical data from the start of sshd to the demise of it and see if there is any choking happens on any aspect of the system.