Re: [PATCH] [NET] Size listen hash tables using backlog hint

[Eric Dumazet <dada1@cosmosbay.com>]

David Miller a écrit :
> From: Eric Dumazet Hi <dada1@cosmosbay.com>
> Date: Tue, 17 Oct 2006 14:58:37 +0200
> 
>> reqsk_queue_alloc() goal is to use a power of two size for the whole
>> listen_sock structure, to avoid wasting memory for large backlogs,
>> meaning the hash table nr_table_entries is not anymore a power of
>> two. (Hence one AND (nr_table_entries - 1) must be replaced by
>> MODULO nr_table_entries)
> 
> Modulus can be very expensive for some small/slow cpus.  Please round
> down to a power-of-2 instead of up if you think the wastage really
> matters.
> 
> Thanks.

I am not sure I understand your points. Rounding up or down still need the 
modulus. Only the size changes by a two factor. I feel you want me to remove 
the modulus, thats unrelated to rounding.

A 66 MHz 486 can perform 1.000.000 divisions per second. Is it a 'slow' cpu ?

If we stay with a power-of-two, say 2^X hash slots, using (2^X)*sizeof(void*), 
the extra bits added by struct listen_sock will *need* the same amount of 
memory, because of kmalloc() alignment to next power-of-two. That basically 
wastes half of the ram taken by struct listen_sock allocation, unless we add 
yet another pointer to hash table and do two kmallocs(), one for pure 
power-of-two hash table, one for struct listen_sock. If we keep current 
scheme, the current max kmalloc size of 131072 bytes would limit us to 65536 
bytes for the hash table itself, so 8192 slots on 64bits platforms. I was 
expecting to use a 16380 slots hash size instead.

The modulus is done on two places :

inet_csk_search_req() : called from tcp_v4_err()/dccp_v4_err() only after 
checks. Frequency of such events is rather low.

tcp_v4_hnd_req() : called from tcp_v4_do_rcv() for TCP_LISTEN state. Frequency 
of such events is rather low, especially on machines driven by small/slow cpus...

inet_csk_reqsk_queue_hash_add()called from tcp_v4_conn_request() when a new 
connection attempt is stored in hash table.

Thats in normal conditions two modulus done per new tcp/dccp sessions 
establishments. In DOS situation, I doubt the extra cycles will do any difference.

So... what do you prefer :

1) Keep the modulus
2) allocate two blocks of ram (powser-of -two hash size, but one extra 
indirection)
3) waste near half of ram because one block allocated, and power-of-two hash size.

Thank you

Eric