Re: How to compile and use 'nscd'

[이병욱]

Thank you for your replies!

Then how about making `CACHE_PRUNE_INTERVAL` tunable at nscd.conf?
 
I think it is too large considering the current server specification.
 
It seems `CACHE_PRUNE_INTERVAL` first appeared at glibc-2.0.99 which was before 2000. It passed about 20 years.
 
I think this large `CACHE_PRUNE_INTERVAL` can make DNS caching meaningless, especially when using dns round robin.
 
At DNS round robin with CACHE_PRUNE_INTERVAL, it makes unequally distributed traffic to servers. It makes inefficient servers usage.
(I did some experiments for checking this, I posted this https://stackoverflow.com/questions/73181902/could-we-use-dns-round-robin-with-nscds-dns-cache)
 
How about making `CACHE_PRUNE_INTERVAL` tunable at nscd.conf or making it smaller? 

Maybe it mitigates the load of CPU at 2000. But If `CACHE_PRUNE_INTERVAL` become smaller, I think it is not that much burdened at current CPU.
(I didn't do this kind of experiments. But If you think it is needed, I will do it.) 

I think it is more reasonable for user to be able to tune `CACHE_PRUNE_INTERVAL` considering user's machine environment.
 
Sincerely,
Nimdrak.
-----Original Message-----
From: "Carlos O'Donell"<carlos@redhat.com>
To: "이병욱"<nimdrak@naver.com>; "Carlos O'Donell"<carlos@systemhalted.org>;
Cc: "libc-help"<libc-help@sourceware.org>;
Sent: 2022-07-30 (토) 01:41:48 (GMT+09:00)
Subject: Re: How to compile and use 'nscd'
 
On 7/24/22 10:15, 이병욱 via Libc-help wrote:
> Thank you for your reply.
>
> If you know, would you tell me why `CACHE_PRUNE_INTERVAL` is defined
> as 15? It is hard to find the reason anywhere.

I don't think there is a strong justification for 15s.

It could be added as a tunable to nscd.conf.

> If it was intended for mitigating the load of the client server with
> nscd, I think it is a little large, considering the current server
> performance.

It limits the CPU usage on the client by pruning the cache at certain
intervals.

The cache entries themselves, depending on the cache, will have TTL
information that can time out on a shorter interval.

> I found at JVM environment (JVM has its own DNS caching), It seems
> that TTL value close to 0, doesn't have an impact to a client server
> performance, to an extent.
>
> I think if we choose a DNS round robin for distributing the traffic
> load, it is inevitable to assign a DNS ttl close to 0. Otherwise, it
> can make biased traffic to servers.
>
> Although I make a A record TTL value below 15, this small TTL value
> become meaningless because the client make a dns query at least every
> 15 seconds, which is `CACHE_PRUNE_INTERVAL`.

Why does it become meaningless?

The CACHE_PRUNE_INTERVAL is the interval in which the pruning thread looks
for entries with expired lifetime. It is not the interval at which entries
are expired, the actual expiry is determined by TTL values.

> Is there a design philosophy about `15`?

None that we have documented.

--
Cheers,
Carlos.