purpose of the skb head pool

purpose of the skb head pool

[Christoph Hellwig]

[sorry, wrong Cc: the first time]

net/core/skbuf.c has a small per-cpu pool to keep some hot skbufs around
instead of returning them to the system allocator.  But if you loook
at the slab allocator we'll have exactly that same code duplicated in
there (see functions ac_data, __cache_alloc and kmem_cache_alloc in
slab.c).  So is there some other reason why this pool is needed?

Re: purpose of the skb head pool

[Andi Kleen]

On Tue, Apr 29, 2003 at 01:55:06PM +0200, Christoph Hellwig wrote:
> [sorry, wrong Cc: the first time]
> 
> net/core/skbuf.c has a small per-cpu pool to keep some hot skbufs around
> instead of returning them to the system allocator.  But if you loook
> at the slab allocator we'll have exactly that same code duplicated in
> there (see functions ac_data, __cache_alloc and kmem_cache_alloc in
> slab.c).  So is there some other reason why this pool is needed?

The code was added before slab grew an own per cpu allocator.

AFAIK it wasn't removed because the "extreme routing" people
like Jamal and Robert O. still saw small advantages, but it's 
probably worth rebenchmarking.

-Andi

purpose of the skb head pool

[Robert Olsson]

[-- Attachment #1: message body text --]
[-- Type: text/plain, Size: 1105 bytes --]


 Hello!

Christoph Hellwig writes:
 > net/core/skbuf.c has a small per-cpu pool to keep some hot skbufs around
 > instead of returning them to the system allocator.  But if you loook
 > at the slab allocator we'll have exactly that same code duplicated in
 > there (see functions ac_data, __cache_alloc and kmem_cache_alloc in
 > slab.c).  So is there some other reason why this pool is needed?

 Well just happened test without it yesterday... 

 Manfred is working on some improvements of the slab (magazine layer) so I  
 tested this. It seems to do improve performance. I also removed the 
 skb_head_pool for a test run.

 2.6.66 IP. Forwarding of two input simplex flows. eth0->eth1, eth2->eth3 
 Fixed affinity CPU0: eth0, eth3. CPU1: eth1, eth2. Which common for routing
 and should be "worst case" for other use. The test should give a very high 
 load on the packet memory system. As seen at least we don't see any
 improvement from skb_head_pool code. 


 Vanilla 2.5.66              381 kpps
 Magazine                    431 kpps
 Magazine + no skb_head_pool 435 kpps


 Cheers.
						--ro


 

[-- Attachment #2: rem_skb_head_pool.pat --]
[-- Type: application/octet-stream, Size: 3581 bytes --]

--- linux/net/core/skbuff.c.030428	2003-04-01 13:24:14.000000000 +0200
+++ linux/net/core/skbuff.c	2003-04-28 16:53:54.000000000 +0200
@@ -20,7 +20,7 @@
  *		Ray VanTassle	:	Fixed --skb->lock in free
  *		Alan Cox	:	skb_copy copy arp field
  *		Andi Kleen	:	slabified it.
+ *		Robert Olsson	:	Removed skb_head_pool
  *
  *	NOTE:
  *		The __skb_ routines should be called with interrupts
@@ -64,15 +64,8 @@
 #include <asm/uaccess.h>
 #include <asm/system.h>
 
-int sysctl_hot_list_len = 128;
-
 static kmem_cache_t *skbuff_head_cache;
 
-static union {
-	struct sk_buff_head	list;
-	char			pad[SMP_CACHE_BYTES];
-} skb_head_pool[NR_CPUS];
-
 /*
  *	Keep out-of-line to prevent kernel bloat.
  *	__builtin_return_address is not used because it is not always
@@ -110,44 +103,6 @@
 	BUG();
 }
 
-static __inline__ struct sk_buff *skb_head_from_pool(void)
-{
-	struct sk_buff_head *list;
-	struct sk_buff *skb = NULL;
-	unsigned long flags;
-
-	local_irq_save(flags);
-
-	list = &skb_head_pool[smp_processor_id()].list;
-
-	if (skb_queue_len(list))
-		skb = __skb_dequeue(list);
-
-	local_irq_restore(flags);
-	return skb;
-}
-
-static __inline__ void skb_head_to_pool(struct sk_buff *skb)
-{
-	struct sk_buff_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-
-	list = &skb_head_pool[smp_processor_id()].list;
-
-	if (skb_queue_len(list) < sysctl_hot_list_len) {
-		__skb_queue_head(list, skb);
-		local_irq_restore(flags);
-
-		return;
-	}
-
-	local_irq_restore(flags);
-	kmem_cache_free(skbuff_head_cache, skb);
-}
-
-
 /* 	Allocate a new skbuff. We do this ourselves so we can fill in a few
  *	'private' fields and also do memory statistics to find all the
  *	[BEEP] leaks.
@@ -182,13 +137,10 @@
 	}
 
 	/* Get the HEAD */
-	skb = skb_head_from_pool();
-	if (!skb) {
-		skb = kmem_cache_alloc(skbuff_head_cache,
+	skb = kmem_cache_alloc(skbuff_head_cache,
 				       gfp_mask & ~__GFP_DMA);
 		if (!skb)
 			goto out;
-	}
 
 	/* Get the DATA. Size must match skb_add_mtu(). */
 	size = SKB_DATA_ALIGN(size);
@@ -207,7 +159,7 @@
 out:
 	return skb;
 nodata:
-	skb_head_to_pool(skb);
+	kmem_cache_free(skbuff_head_cache, skb);
 	skb = NULL;
 	goto out;
 }
@@ -257,7 +209,7 @@
 void kfree_skbmem(struct sk_buff *skb)
 {
 	skb_release_data(skb);
-	skb_head_to_pool(skb);
+	kmem_cache_free(skbuff_head_cache, skb);
 }
 
 /**
@@ -327,13 +279,10 @@
 
 struct sk_buff *skb_clone(struct sk_buff *skb, int gfp_mask)
 {
-	struct sk_buff *n = skb_head_from_pool();
+	struct sk_buff *n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
 
-	if (!n) {
-		n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
-		if (!n)
-			return NULL;
-	}
+	if (!n) 
+		return NULL;
 
 #define C(x) n->x = skb->x
 
@@ -1240,7 +1189,4 @@
 					      NULL, NULL);
 	if (!skbuff_head_cache)
 		panic("cannot create skbuff cache");
-
-	for (i = 0; i < NR_CPUS; i++)
-		skb_queue_head_init(&skb_head_pool[i].list);
 }
--- linux/net/core/sysctl_net_core.c.030428	2003-03-24 23:00:18.000000000 +0100
+++ linux/net/core/sysctl_net_core.c	2003-04-28 16:59:05.000000000 +0200
@@ -28,7 +28,6 @@
 
 extern int sysctl_core_destroy_delay;
 extern int sysctl_optmem_max;
-extern int sysctl_hot_list_len;
 
 #ifdef CONFIG_NET_DIVERT
 extern char sysctl_divert_version[];
@@ -150,14 +149,6 @@
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec
 	},
-	{
-		.ctl_name	= NET_CORE_HOT_LIST_LENGTH,
-		.procname	= "hot_list_length",
-		.data		= &sysctl_hot_list_len,
-		.maxlen		= sizeof(int),
-		.mode		= 0644,
-		.proc_handler	= &proc_dointvec
-	},
 #ifdef CONFIG_NET_DIVERT
 	{
 		.ctl_name	= NET_CORE_DIVERT_VERSION,

[-- Attachment #3: message body text --]
[-- Type: text/plain, Size: 8542 bytes --]




 Vanilla slab. Input streams 2*534 kpps
======================================

           CPU0       CPU1       
 24:          9         65   IO-APIC-level  eth2
 25:      54545         13   IO-APIC-level  eth3
 26:         78          0   IO-APIC-level  eth0
 27:         23      62315   IO-APIC-level  eth1

Iface   MTU Met  RX-OK RX-ERR RX-DRP RX-OVR  TX-OK TX-ERR TX-DRP TX-OVR Flags
eth0   1500   0 3339339 8828613 8828613 6660669     33      0      0      0 BRU
eth1   1500   0     57      0      0      0 3339340      0      0      0 BRU
eth2   1500   0 3800145 8670213 8670213 6199858     27      0      0      0 BRU
eth3   1500   0      1      0      0      0 3800144      0      0      0 BRU

0032f44f 00000000 00002ffa 00000000 00000000 00000000 00000000 00000000 00000000
0039fc87 00000000 00003373 00000000 00000000 00000000 00000000 00000000 00000000

With slab magazine patch. Input streams 2*534 kpps
==================================================

Iface   MTU Met  RX-OK RX-ERR RX-DRP RX-OVR  TX-OK TX-ERR TX-DRP TX-OVR Flags
eth0   1500   0 3936399 8370562 8370562 6063606     31      0      0      0 BRU
eth1   1500   0     58      0      0      0 3936403      0      0      0 BRU
eth2   1500   0 4142687 8308862 8308862 5857316     27      0      0      0 BRU
eth3   1500   0      1      0      0      0 4142686      0      0      0 BRU

003c1090 00000000 000034d1 00000000 00000000 00000000 00000000 00000000 00000000
003f3699 00000000 00003722 00000000 00000000 00000000 00000000 00000000 00000000

           CPU0       CPU1       
 24:          9         81   IO-APIC-level  eth2
 25:      64461         14   IO-APIC-level  eth3
 26:         94          0   IO-APIC-level  eth0
 27:         20      67759   IO-APIC-level  eth1
Daemon started.
Profiler running.
Stopping profiling.
Cpu type: P4 / Xeon
Cpu speed was (MHz estimation) : 1799.59
Counter 0 counted GLOBAL_POWER_EVENTS events (time during which processor is not stopped) with a unit mask of 0x01 (count cycles when processor is active) count 180000
vma      samples  %-age       symbol name
c02152c0 89894    13.7456     alloc_skb
c022af44 69317    10.5992     ip_output
c021fd90 64410    9.84889     qdisc_restart
c01c8608 60224    9.20882     e1000_clean_tx_irq
c021fbb8 48501    7.41626     eth_type_trans
c021558c 42143    6.44406     skb_release_data
c02156e4 36162    5.52951     __kfree_skb
c01378b8 26285    4.01922     kmalloc
c022645c 25858    3.95393     ip_route_input
c01c87c0 23820    3.6423      e1000_clean_rx_irq
c01c76c8 23261    3.55683     e1000_xmit_frame
c0218a60 17677    2.70298     dev_queue_xmit
c01374e0 12685    1.93966     cache_alloc_refill
c0137a00 11319    1.73078     kfree
c010fda0 11077    1.69378     do_gettimeofday
c0228314 10202    1.55998     ip_rcv
c0114050 8899     1.36074     get_offset_tsc
c021567c 8394     1.28352     kfree_skbmem
c0229700 6794     1.03887     ip_forward
c01c8bf0 6531     0.998651    e1000_alloc_rx_buffers
c01c8554 6379     0.975409    e1000_clean
c021cadc 5234     0.800328    neigh_resolve_output
c02190f0 5106     0.780755    netif_receive_skb
c01c8468 4658     0.712252    e1000_intr
c0114068 3542     0.541605    mark_offset_tsc
c0222320 2156     0.329673    pfifo_dequeue
Cpu type: P4 / Xeon
Cpu speed was (MHz estimation) : 1799.59
Counter 7 counted MISPRED_BRANCH_RETIRED events (retired mispredicted branches) with a unit mask of 0x01 (retired instruction is non-bogus) count 18000
vma      samples  %-age       symbol name
c022af44 718      27.7113     ip_output
c021fd90 513      19.7993     qdisc_restart
c02156e4 360      13.8942     __kfree_skb
c01c8608 261      10.0733     e1000_clean_tx_irq
c0218a60 85       3.28059     dev_queue_xmit
c01c8bf0 85       3.28059     e1000_alloc_rx_buffers
c0137a00 53       2.04554     kfree
c01c87c0 50       1.92976     e1000_clean_rx_irq
c021558c 48       1.85257     skb_release_data
c01378b8 46       1.77538     kmalloc
c021fbb8 41       1.5824      eth_type_trans
c0222320 36       1.38942     pfifo_dequeue
c0228314 31       1.19645     ip_rcv
c01c76c8 30       1.15785     e1000_xmit_frame
c02152c0 24       0.926283    alloc_skb
c022645c 19       0.733308    ip_route_input
c01377d0 19       0.733308    cache_flusharray
c010c750 17       0.656117    do_IRQ
c01c8554 15       0.578927    e1000_clean
c01168b8 13       0.501737    end_level_ioapic_irq
c02190f0 12       0.463142    netif_receive_skb
c01374e0 12       0.463142    cache_alloc_refill
c0120eb0 11       0.424547    do_softirq
c0229700 10       0.385951    ip_forward
c01c8468 10       0.385951    e1000_intr
c01245e8 9        0.347356    run_timer_softirq


With slab magazine patch and skb_head_pool removed. Input streams 2*533 kpps
============================================================================

Iface   MTU Met  RX-OK RX-ERR RX-DRP RX-OVR  TX-OK TX-ERR TX-DRP TX-OVR Flags
eth0   1500   0 4070842 8257568 8257568 5929162     32      0      0      0 BRU
eth1   1500   0     60      0      0      0 4070844      0      0      0 BRU
eth2   1500   0 4097594 8285413 8285413 5902409     27      0      0      0 BRU
eth3   1500   0      1      0      0      0 4097593      0      0      0 BRU

003e1dc4 00000000 000036e4 00000000 00000000 00000000 00000000 00000000 00000000
003e866f 00000000 00003711 00000000 00000000 00000000 00000000 00000000 00000000

           CPU0       CPU1       
 24:          9        156   IO-APIC-level  eth2
 25:      66506          8   IO-APIC-level  eth3
 26:        170          0   IO-APIC-level  eth0
 27:         23      66807   IO-APIC-level  eth1
NMI:          0          0 
LOC:     357638     357637 
ERR:          0
MIS:          0

Daemon started.
Profiler running.
Stopping profiling.
Cpu type: P4 / Xeon
Cpu speed was (MHz estimation) : 1799.55
Counter 0 counted GLOBAL_POWER_EVENTS events (time during which processor is not stopped) with a unit mask of 0x01 (count cycles when processor is active) count 180000
vma      samples  %-age       symbol name
c02152c0 90698    12.1819     alloc_skb
c022adf4 81812    10.9884     ip_output
c021fc40 74295    9.97875     qdisc_restart
c01c8608 67518    9.06852     e1000_clean_tx_irq
c021fa68 55350    7.4342      eth_type_trans
c02154ec 45940    6.17032     skb_release_data
c02155f8 41834    5.61883     __kfree_skb
c01c87c0 31218    4.19297     e1000_clean_rx_irq
c022630c 29209    3.92314     ip_route_input
c01c76c8 27768    3.72959     e1000_xmit_frame
c01378b8 20298    2.72628     kmalloc
c0218910 19762    2.65428     dev_queue_xmit
c01374e0 14245    1.91328     cache_alloc_refill
c0137a00 13807    1.85445     kfree
c010fda0 13406    1.80059     do_gettimeofday
c0137874 12601    1.69247     kmem_cache_alloc
c02281c4 11403    1.53157     ip_rcv
c0114050 9854     1.32352     get_offset_tsc
c01379b8 8743     1.17429     kmem_cache_free
c01c8bf0 8369     1.12406     e1000_alloc_rx_buffers
c02295b0 7607     1.02172     ip_forward
c01c8554 6912     0.928368    e1000_clean
c01c8468 5636     0.756986    e1000_intr
c0218fa0 5413     0.727034    netif_receive_skb
c01d2194 4363     0.586006    ide_insw
c0114068 3972     0.533489    mark_offset_tsc
Cpu type: P4 / Xeon
Cpu speed was (MHz estimation) : 1799.55
Counter 7 counted MISPRED_BRANCH_RETIRED events (retired mispredicted branches) with a unit mask of 0x01 (retired instruction is non-bogus) count 18000
vma      samples  %-age       symbol name
c022adf4 766      26.3049     ip_output
c021fc40 617      21.1882     qdisc_restart
c02155f8 378      12.9808     __kfree_skb
c01c8608 298      10.2335     e1000_clean_tx_irq
c01c8bf0 120      4.12088     e1000_alloc_rx_buffers
c0218910 114      3.91484     dev_queue_xmit
c01c87c0 85       2.91896     e1000_clean_rx_irq
c0137a00 71       2.43819     kfree
c021fa68 42       1.44231     eth_type_trans
c02154ec 40       1.37363     skb_release_data
c01c76c8 39       1.33929     e1000_xmit_frame
c01378b8 31       1.06456     kmalloc
c02221d0 27       0.927198    pfifo_dequeue
c02295b0 23       0.789835    ip_forward
c02281c4 22       0.755495    ip_rcv
c02152c0 20       0.686813    alloc_skb
c01c8468 17       0.583791    e1000_intr
c01377d0 17       0.583791    cache_flusharray
c01c8554 16       0.549451    e1000_clean
c022630c 15       0.51511     ip_route_input
c01374e0 14       0.480769    cache_alloc_refill
c01168b8 13       0.446429    end_level_ioapic_irq
c010c750 10       0.343407    do_IRQ
c0114050 9        0.309066    get_offset_tsc
c0110074 9        0.309066    timer_interrupt
c02191f4 8        0.274725    net_rx_action
 

Re: purpose of the skb head pool

[David S. Miller]

   From: Robert Olsson <Robert.Olsson@data.slu.se>
   Date: Tue, 29 Apr 2003 15:11:38 +0200

   Do we need a test run without Manfreds (soon coming) magazine patch too?

Nope.  I'm personally convinced, someone send me a patch to
kill it off and I'll apply it.

Re: purpose of the skb head pool

[Robert Olsson]

Andi Kleen writes:

 > AFAIK it wasn't removed because the "extreme routing" people
 > like Jamal and Robert O. still saw small advantages, but it's 
 > probably worth rebenchmarking.

 :-)
 Do we need a test run without Manfreds (soon coming) magazine patch too?

 Cheers.

						--ro

Re: purpose of the skb head pool

[David S. Miller]

   From: Christoph Hellwig <hch@lst.de>
   Date: Tue, 29 Apr 2003 16:16:08 +0200

   On Tue, Apr 29, 2003 at 06:08:34AM -0700, David S. Miller wrote:
   > Nope.  I'm personally convinced, someone send me a patch to
   > kill it off and I'll apply it.
   
   Robert's first reply on this subject had it attached..
   
I deleted it, can someone resend me a copy?

When I ask for a patch, it usually means that I've
/dev/null'd the entire thread already.

Re: purpose of the skb head pool

[Christoph Hellwig]

On Tue, Apr 29, 2003 at 06:08:34AM -0700, David S. Miller wrote:
>    From: Robert Olsson <Robert.Olsson@data.slu.se>
>    Date: Tue, 29 Apr 2003 15:11:38 +0200
> 
>    Do we need a test run without Manfreds (soon coming) magazine patch too?
> 
> Nope.  I'm personally convinced, someone send me a patch to
> kill it off and I'll apply it.

Robert's first reply on this subject had it attached..

Re: purpose of the skb head pool

[Christoph Hellwig]

On Tue, Apr 29, 2003 at 06:12:49AM -0700, David S. Miller wrote:
>    Robert's first reply on this subject had it attached..
>    
> I deleted it, can someone resend me a copy?
> 
> When I ask for a patch, it usually means that I've
> /dev/null'd the entire thread already.

Lazy bastard.. ;)


--- linux/net/core/skbuff.c.030428	2003-04-01 13:24:14.000000000 +0200
+++ linux/net/core/skbuff.c	2003-04-28 16:53:54.000000000 +0200
@@ -20,7 +20,7 @@
  *		Ray VanTassle	:	Fixed --skb->lock in free
  *		Alan Cox	:	skb_copy copy arp field
  *		Andi Kleen	:	slabified it.
+ *		Robert Olsson	:	Removed skb_head_pool
  *
  *	NOTE:
  *		The __skb_ routines should be called with interrupts
@@ -64,15 +64,8 @@
 #include <asm/uaccess.h>
 #include <asm/system.h>
 
-int sysctl_hot_list_len = 128;
-
 static kmem_cache_t *skbuff_head_cache;
 
-static union {
-	struct sk_buff_head	list;
-	char			pad[SMP_CACHE_BYTES];
-} skb_head_pool[NR_CPUS];
-
 /*
  *	Keep out-of-line to prevent kernel bloat.
  *	__builtin_return_address is not used because it is not always
@@ -110,44 +103,6 @@
 	BUG();
 }
 
-static __inline__ struct sk_buff *skb_head_from_pool(void)
-{
-	struct sk_buff_head *list;
-	struct sk_buff *skb = NULL;
-	unsigned long flags;
-
-	local_irq_save(flags);
-
-	list = &skb_head_pool[smp_processor_id()].list;
-
-	if (skb_queue_len(list))
-		skb = __skb_dequeue(list);
-
-	local_irq_restore(flags);
-	return skb;
-}
-
-static __inline__ void skb_head_to_pool(struct sk_buff *skb)
-{
-	struct sk_buff_head *list;
-	unsigned long flags;
-
-	local_irq_save(flags);
-
-	list = &skb_head_pool[smp_processor_id()].list;
-
-	if (skb_queue_len(list) < sysctl_hot_list_len) {
-		__skb_queue_head(list, skb);
-		local_irq_restore(flags);
-
-		return;
-	}
-
-	local_irq_restore(flags);
-	kmem_cache_free(skbuff_head_cache, skb);
-}
-
-
 /* 	Allocate a new skbuff. We do this ourselves so we can fill in a few
  *	'private' fields and also do memory statistics to find all the
  *	[BEEP] leaks.
@@ -182,13 +137,10 @@
 	}
 
 	/* Get the HEAD */
-	skb = skb_head_from_pool();
-	if (!skb) {
-		skb = kmem_cache_alloc(skbuff_head_cache,
+	skb = kmem_cache_alloc(skbuff_head_cache,
 				       gfp_mask & ~__GFP_DMA);
 		if (!skb)
 			goto out;
-	}
 
 	/* Get the DATA. Size must match skb_add_mtu(). */
 	size = SKB_DATA_ALIGN(size);
@@ -207,7 +159,7 @@
 out:
 	return skb;
 nodata:
-	skb_head_to_pool(skb);
+	kmem_cache_free(skbuff_head_cache, skb);
 	skb = NULL;
 	goto out;
 }
@@ -257,7 +209,7 @@
 void kfree_skbmem(struct sk_buff *skb)
 {
 	skb_release_data(skb);
-	skb_head_to_pool(skb);
+	kmem_cache_free(skbuff_head_cache, skb);
 }
 
 /**
@@ -327,13 +279,10 @@
 
 struct sk_buff *skb_clone(struct sk_buff *skb, int gfp_mask)
 {
-	struct sk_buff *n = skb_head_from_pool();
+	struct sk_buff *n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
 
-	if (!n) {
-		n = kmem_cache_alloc(skbuff_head_cache, gfp_mask);
-		if (!n)
-			return NULL;
-	}
+	if (!n) 
+		return NULL;
 
 #define C(x) n->x = skb->x
 
@@ -1240,7 +1189,4 @@
 					      NULL, NULL);
 	if (!skbuff_head_cache)
 		panic("cannot create skbuff cache");
-
-	for (i = 0; i < NR_CPUS; i++)
-		skb_queue_head_init(&skb_head_pool[i].list);
 }
--- linux/net/core/sysctl_net_core.c.030428	2003-03-24 23:00:18.000000000 +0100
+++ linux/net/core/sysctl_net_core.c	2003-04-28 16:59:05.000000000 +0200
@@ -28,7 +28,6 @@
 
 extern int sysctl_core_destroy_delay;
 extern int sysctl_optmem_max;
-extern int sysctl_hot_list_len;
 
 #ifdef CONFIG_NET_DIVERT
 extern char sysctl_divert_version[];
@@ -150,14 +149,6 @@
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec
 	},
-	{
-		.ctl_name	= NET_CORE_HOT_LIST_LENGTH,
-		.procname	= "hot_list_length",
-		.data		= &sysctl_hot_list_len,
-		.maxlen		= sizeof(int),
-		.mode		= 0644,
-		.proc_handler	= &proc_dointvec
-	},
 #ifdef CONFIG_NET_DIVERT
 	{
 		.ctl_name	= NET_CORE_DIVERT_VERSION,

Re: purpose of the skb head pool

[David S. Miller]

   From: Christoph Hellwig <hch@lst.de>
   Date: Tue, 29 Apr 2003 16:21:02 +0200

   On Tue, Apr 29, 2003 at 06:12:49AM -0700, David S. Miller wrote:
   >    Robert's first reply on this subject had it attached..
   >    
   > I deleted it, can someone resend me a copy?
   > 
   > When I ask for a patch, it usually means that I've
   > /dev/null'd the entire thread already.
   
   Lazy bastard.. ;)
   
Hehe :-)  Applied, thanks.

Re: purpose of the skb head pool

[David S. Miller]

   From: Florian Weimer <fw@deneb.enyo.de>
   Date: Thu, 01 May 2003 12:38:38 +0200

   Robert Olsson <Robert.Olsson@data.slu.se> writes:
   
   >  Vanilla 2.5.66              381 kpps
   >  Magazine                    431 kpps
   >  Magazine + no skb_head_pool 435 kpps
   
   Can you rerun this test with random source/destination addresses, to
   get more realistic (for some configurations) numbers?
   
He can do this, but the issues we're trying to tackle first have
nothing to even do with what kind of routing cache accesses
are done.

It's all networking buffer overhead we're worried about at this
stage.

Re: purpose of the skb head pool

[Florian Weimer]

Robert Olsson <Robert.Olsson@data.slu.se> writes:

>  2.6.66 IP. Forwarding of two input simplex flows. eth0->eth1, eth2->eth3 
>  Fixed affinity CPU0: eth0, eth3. CPU1: eth1, eth2. Which common for routing
>  and should be "worst case" for other use. The test should give a very high 
>  load on the packet memory system. As seen at least we don't see any
>  improvement from skb_head_pool code. 
>
>
>  Vanilla 2.5.66              381 kpps
>  Magazine                    431 kpps
>  Magazine + no skb_head_pool 435 kpps

Can you rerun this test with random source/destination addresses, to
get more realistic (for some configurations) numbers?