128 Bucket Failures?

[Chris Pratt]

On Nov 13, 2008, at 1:34 PM, Ivan Voras wrote:

> Chris Pratt wrote:
>> I have asked this before a couple of years ago but received no
>> replies. I assumed that's because it's a somewhat obscure question.
>> I'm still interested and thought I might try again in case someone
>> new is watching this list who might know.
>>
>> A vmstat -z on my highest traffic server always shows the failures
>> as below on 128 Bucket. It also goes to having 0 free rather soon
>> after the system is restarted and never returns to having more than
>> 1 free in that column and yet always has the highest number of
>> requests by far. Does this mean anything significant? Is it
>> something I should tune or even can be tuned?
>
> UMA buckets seem to be some kind of cache for SMP-optimized  
> allocations
> - I hope someone who knows it better will explain them.
>
>> Here is the output of the vmstat -z with everything chopped out
>> besides the 128 Bucket line. The machine it's on is an 8 core 8 GB
>> Tyan and shouldn't really be starved for anything in my way of  
>> thinking.
>>
>> vmstat -z
>> ITEM                     SIZE     LIMIT      USED      FREE   
>> REQUESTS  FAILURES
>>
>> 128 Bucket:              1048,        0,     2043,        0,     
>> 13591,  6511069
>
> What is the server used for?
>

A busy webserver (about 5G Views a month, average view is 3-4 hits).  
Not really
large pages, we keep graphics minimal. It's apache, perl cgi, mysqld.  
Tends to
collect a lot of garbage traffic attacks on top of real traffic, both  
TCP and UDP.

> Here's a snapshot from a very loaded apache+php+pgsql web server,  
> uptime
> 60 days (since the last power outage):
>
> 16 Bucket:                 76,        0,       42,       58,      125,
>       0
> 32 Bucket:                140,        0,       76,       64,      183,
>       0
> 64 Bucket:                268,        0,       74,       38,      438,
>      11
> 128 Bucket:               524,        0,     2060,      642,   788828,
>    6985
>
> A generic advice would be to increase vm.kmem_size (you're using  
> AMD64,
> right?) and see what happens.
>

I'll try that. I had heard this before in relation to KVA but have  
been concerned
about trying it. If I can just change that knob and have an effect,  
seems worth
a try. If more than one person is doing it, it must be safe?

Yes, AMD64. Thank you very much.