Re: (RFC): SKB Initialization

Re: (RFC): SKB Initialization

[Mala Anand]

Here is a 2.5.25 patch that improves the skb header initialization
on SMP systems [for netperf].  The patch removes the constructor
for the skbuff_head_cache and moves the initialization of the skb
header from __free_skb to alloc_skb.

Since skbs freed on one CPU may end up being allocated on another cpu,
and often are, the initialization on the free side may not be the
right choice for SMP particularily if the free slab is then allocated
on another CPU.

I measured the cycles spent in initializing skb in the stock kernel
versus with the patch. I collected the cycles at the end of running
tcp_stream netperf test using 4KB message size, 64KB socket buffer
size, 2 adapter test on a 2-way system.

Two 996MHz Pentium III systems with 2 gigabit ethernet cards in each
one of them are used. The cycles are collected at the server
end.

The patch reduces the numer of cylces by 25%

Baseline on Linux 2.5.25 kernel:
-------------------------------

                                CPU 0                 CPU 1
                               ------                 ------
Avg cycles in alloc_skb:        64.05                 203.39
Avg cycles in __kfree_skb:     127.54                 228.95
                               ------                 -------
Total Avg Cycles               191.59                 432.34
                               ------                 -------

# of times alloc_skb called:      235,478            2,060,422
# of times __kfree_skb called:  2,063,276              232,359


Linux 2.5.25+Skbinit Patch:
--------------------------
                              CPU 0                   CPU 1
                              -----                   -----
 Avg cycles in alloc skb:     237.21                  230.91

 # of times alloc_skb called: 1,226,594             1,213,327



I also will collect the cycles of this init code in Specweb99 workload
and will post it later.



diff -Naur linux2.5.25/net/core/skbuff.c linux-525skb/net/core/skbuff.c
--- linux2.5.25/net/core/skbuff.c   Tue Aug 20 08:50:58 2002
+++ linux-525skb/net/core/skbuff.c  Wed Aug 21 08:51:52 2002
@@ -195,18 +195,37 @@
            goto nodata;

      /* XXX: does not include slab overhead */
+     skb->next     = skb->prev = NULL;
+     skb->list     = NULL;
+     skb->sk             = NULL;
+     skb->stamp.tv_sec = 0;  /* No idea about time */
+     skb->dev      = NULL;
+     skb->dst      = NULL;
+     memset(skb->cb, 0, sizeof(skb->cb));
+     skb->len      = 0;
+     skb->data_len = 0;
+     skb->csum = 0;
+     skb->cloned   = 0;
+     skb->pkt_type       = PACKET_HOST;  /* Default type */
+     skb->ip_summed      = 0;
+     skb->priority       = 0;
+     skb->security       = 0;      /* By default packets are insecure */
      skb->truesize = size + sizeof(struct sk_buff);
-
      /* Load the data pointers. */
      skb->head = skb->data = skb->tail = data;
      skb->end  = data + size;
-
-     /* Set up other state */
-     skb->len      = 0;
-     skb->cloned   = 0;
-     skb->data_len = 0;
-
      atomic_set(&skb->users, 1);
+     skb->destructor     = NULL;
+#ifdef CONFIG_NETFILTER
+     skb->nfmark   = skb->nfcache = 0;
+     skb->nfct     = NULL;
+#ifdef CONFIG_NETFILTER_DEBUG
+     skb->nf_debug       = 0;
+#endif
+#endif
+#ifdef CONFIG_NET_SCHED
+     skb->tc_index       = 0;
+#endif
      atomic_set(&(skb_shinfo(skb)->dataref), 1);
      skb_shinfo(skb)->nr_frags  = 0;
      skb_shinfo(skb)->frag_list = NULL;
@@ -326,7 +345,7 @@
 #ifdef CONFIG_NETFILTER
      nf_conntrack_put(skb->nfct);
 #endif
-     skb_headerinit(skb, NULL, 0);  /* clean state */
+     /* skb_headerinit(skb, NULL, 0);   clean state */
      kfree_skbmem(skb);
 }

@@ -1191,7 +1210,7 @@
                                    sizeof(struct sk_buff),
                                    0,
                                    SLAB_HWCACHE_ALIGN,
-                                   skb_headerinit, NULL);
+                                   NULL, NULL);
      if (!skbuff_head_cache)
            panic("cannot create skbuff cache");




Regards,
    Mala


   Mala Anand
   IBM Linux Technology Center - Kernel Performance
   E-mail:manand@us.ibm.com
   http://www-124.ibm.comdeveloperworks/opensource/linuxperf
   http://www-124.ibm.com/developerworks/projects/linuxperf
   Phone:838-8088; Tie-line:678-8088

Re: (RFC): SKB Initialization

[Benjamin LaHaise]

On Wed, Aug 21, 2002 at 11:59:44AM -0500, Mala Anand wrote:
> The patch reduces the numer of cylces by 25%

The data you are reporting is flawed: where are the average cycle 
times spent in __kfree_skb with the patch?  Without that information, 
there is no basis for conclusion as the total time spent in these 
two routines is unknown.  At worst, I'd have to assume this lack of 
scientific method is an attempt to hide some aspect of the resulting 
behaviour, but I hope that isn't the case.

		-ben

> Baseline on Linux 2.5.25 kernel:
> -------------------------------
> 
>                                 CPU 0                 CPU 1
>                                ------                 ------
> Avg cycles in alloc_skb:        64.05                 203.39
> Avg cycles in __kfree_skb:     127.54                 228.95
>                                ------                 -------
> Total Avg Cycles               191.59                 432.34
>                                ------                 -------
> 
> # of times alloc_skb called:      235,478            2,060,422
> # of times __kfree_skb called:  2,063,276              232,359
> 
> 
> Linux 2.5.25+Skbinit Patch:
> --------------------------
>                               CPU 0                   CPU 1
>                               -----                   -----
>  Avg cycles in alloc skb:     237.21                  230.91
> 
>  # of times alloc_skb called: 1,226,594             1,213,327

Re: (RFC): SKB Initialization

[Mala Anand]

>On Wed, Aug 21, 2002 at 11:59:44AM -0500, Mala Anand wrote:
>> The patch reduces the numer of cylces by 25%

>The data you are reporting is flawed: where are the average cycle
>times spent in __kfree_skb with the patch?

I measured the cycles for only the initialization code in alloc_skb
and __kfree_skb. Since the init code is removed from __kfree_skb,
no cycles are spent there.



Regards,
    Mala


   Mala Anand
   IBM Linux Technology Center - Kernel Performance
   E-mail:manand@us.ibm.com
   http://www-124.ibm.comdeveloperworks/opensource/linuxperf
   http://www-124.ibm.com/developerworks/projects/linuxperf
   Phone:838-8088; Tie-line:678-8088

Re: (RFC): SKB Initialization

[Benjamin LaHaise]

On Wed, Aug 21, 2002 at 01:07:09PM -0500, Mala Anand wrote:
> 
> >On Wed, Aug 21, 2002 at 11:59:44AM -0500, Mala Anand wrote:
> >> The patch reduces the numer of cylces by 25%
> 
> >The data you are reporting is flawed: where are the average cycle
> >times spent in __kfree_skb with the patch?
> 
> I measured the cycles for only the initialization code in alloc_skb
> and __kfree_skb. Since the init code is removed from __kfree_skb,
> no cycles are spent there.

Then the testing technique is flawed.  You should include all of the 
operations included in an alloc_skb/kfree_skb pair in order to see 
the overall effect of the change, otherwise your change could have a 
net negative effect which would not be noticed.

		-ben
-- 
"You will be reincarnated as a toad; and you will be much happier."

Re: (RFC): SKB Initialization

[Nivedita Singhvi]

> The patch reduces the numer of cylces by 25%

Er, I'm not getting that from just the numbers you show.

The total ave number of cycles in alloc_skb before is approx 268.

Total number of cycles in alloc_skb after is approx 468. 

So in order to reduce the total ave number of cycles, you
need __kfree_skb to reduce by more than approx 200. (Since
thats where you moved the code to).

So unless we see the total cycles in __kfree_skb, how
do we know? Or what am I missing?

thanks,
Nivedita


You said:



                                CPU 0                 CPU 1
                               ------                 ------
Avg cycles in alloc_skb:        64.05                 203.39
Avg cycles in __kfree_skb:     127.54                 228.95
                               ------                 -------
Total Avg Cycles               191.59                 432.34
                               ------                 -------

# of times alloc_skb called:      235,478            2,060,422
# of times __kfree_skb called:  2,063,276              232,359


Linux 2.5.25+Skbinit Patch:
--------------------------
                              CPU 0                   CPU 1
                              -----                   -----
 Avg cycles in alloc skb:     237.21                  230.91

 # of times alloc_skb called: 1,226,594             1,213,327