Re: [RFC][PATCH 1/6] Storing ipcs into radix trees

[Nadia Derbey <Nadia.Derbey@bull.net>]

Ingo Oeser wrote:
> Hi Nadia,
> 
> good to see someone is pounding this old beast again :-)
> 
> On Thursday 07 June 2007, Nadia.Derbey@bull.net wrote:
> 
>>Index: linux-2.6.21/ipc/util.h
>>===================================================================
>>--- linux-2.6.21.orig/ipc/util.h	2007-06-07 11:00:30.000000000 +0200
>>+++ linux-2.6.21/ipc/util.h	2007-06-07 11:07:22.000000000 +0200
>>@@ -13,6 +13,8 @@
>> #define USHRT_MAX 0xffff
>> #define SEQ_MULTIPLIER	(IPCMNI)
>> 
>>+#define IPCS_MAX_SCAN_ENTRIES 256
> 
> 
> That ...
> 
> 
>>Index: linux-2.6.21/ipc/util.c
>>===================================================================
>>--- linux-2.6.21.orig/ipc/util.c	2007-06-07 11:00:30.000000000 +0200
>>+++ linux-2.6.21/ipc/util.c	2007-06-07 11:29:43.000000000 +0200
>>@@ -252,72 +241,94 @@ void __init ipc_init_proc_interface(cons
>>  *	@key: The key to find
>>  *	
>>  *	Requires ipc_ids.mutex locked.
>>- *	Returns the identifier if found or -1 if not.
>>+ *	Returns the LOCKED pointer to the ipc structure if found or NULL
>>+ *	if not.
>>+ *      If key is found ipc contains its ipc structure
>>  */
>>  
>>-int ipc_findkey(struct ipc_ids* ids, key_t key)
>>+struct kern_ipc_perm *ipc_findkey(struct ipc_ids *ids, key_t key)
>> {
>>-	int id;
>>-	struct kern_ipc_perm* p;
>>-	int max_id = ids->max_id;
>>+	struct kern_ipc_perm *ipc;
>>+	struct kern_ipc_perm *ipcs[IPCS_MAX_SCAN_ENTRIES];
> 
> 
> ... together with this means 4*256 -> 1k of precious stack space used.
> Please consider either lowering IPCS_MAX_SCAN_ENTRIES or kmalloc() that.
> 
> Same problem with your third patch called 
> "Changing the loops on a single ipcid into radix_tree_gang_lookup() calls"
> 
> If you cannot sleep, try to lower that constant (e.g. 16-32). 
> The current users use much smaller numbers.
> 
> If you can sleep and performance goes down after lowering that constant,
> try to kmalloc these arrays (since kmalloc() of that small amount 
> should succeed easily).
> 
> 
> 
> Regards
> 
> Ingo Oeser
> 

Ingo,

You're completely right, but trying to lower the extraction size, I'm 
afraid this will have an impact on performances.

Here are the results of a small test I did: I have run ctxbench on both 
the 256 and and 16 entries versions

1) 256 entries:

[root@akt ctxbench-1.9]# echo 1000 > /proc/sys/kernel/msgmni
[root@akt ctxbench-1.9]# ./ctx -m -t300


Context switching benchmark v1.17
   Using message queue for IPC control
    Max iterations:        0 (zero = no limit)
Max runtime (sec):      300 (zero = no limit)
42523679 itterations in 300.005423 seconds = 141743/sec

2) 16 entries:

[root@akt ctxbench-1.9]# echo 1000 > /proc/sys/kernel/msgmni
[root@akt ctxbench-1.9]# ./ctx -m -t300


Context switching benchmark v1.17
   Using message queue for IPC control
    Max iterations:        0 (zero = no limit)
Max runtime (sec):      300 (zero = no limit)
41774255 itterations in 300.005334 seconds = 139245/sec

Will try with a dynamic allocation.

Regards,
Nadia