Re: Is this skb recycle buffer helpful to improve Linux network stack performance?

Here is the modifications I made:
//skbuff.h
//I added the folloing definitions
#define MAX_POOL_SIZE 4096
unsigned char *skbuff_data_pool[MAX_POOL_SIZE];
int skbPoolHead, skbPoolTail;


//skb_init()
//I add the following codes:
 for(i=0;i<MAX_POOL_SIZE;i++)
         skbuff_data_pool[i] = NULL;
 skbPoolHead = skbPoolTail = 0;


//alloc_skb()
//I made the following changes
if (skbuff_data_pool[skbPoolHead])
 {
   data = skbuff_data_pool[skbPoolHead];
  skbuff_data_pool[skbPoolHead] = NULL;
  if (++skbPoolHead == MAX_POOL_SIZE)
   skbPoolHead = 0;
 }
 else{ //Original path
  size = SKB_DATA_ALIGN(size);
  data = kmalloc(size + sizeof(struct skb_shared_info), gfp_mask);
  }

//skb_release_data()
//I made the following changes:
  if ((skbPoolHead == skbPoolTail) && (skbuff_data_pool[skbPoolHead] != 
NULL))
  //Original path
   kfree(skb->head);
                else{
   if (skbuff_data_pool[skbPoolTail])
    panic("Tail pointer must be null");

   skbuff_data_pool[skbPoolTail] = skb->head;
                        if (++skbPoolTail == MAX_POOL_SIZE)
    skbPoolTail = 0;
  }

----- Original Message ----- 
From: "Eric Dumazet" <redacted>
To: "LeoY" <redacted>
Cc: <redacted>
Sent: Saturday, October 08, 2005 2:47 PM
Subject: Re: Is this skb recycle buffer helpful to improve Linux network 
stack performance?

LeoY a écrit :

quoted

Hi,

Motivation: we noticed alloc_skb()/kfree() used lots of clock ticks when
handling heavy network traffic. As Linux kernel always need to call
kmalloc()/kfree() to allocate and deallocate a skb DATA buffer(not 
sk_buff)
for each incoming/outgoing packet, we try to reduce the frequence of 
calling
these memory functions.

I wangt to set up a ring buffer in Linux kernel(skbuff.c) and recycle 
those
skb data buffers. The basic idea is as follows:
1. Create a ring buffer. This ring buffer has a head pointer which points 
to
the virtual address of the data buffer to be reused; It also has a tail
pointer, which can be used to store the virutal address of skb data 
buffer
for those transmitted packets.
2. If the ring buffer is full, just use normal kmalloc()/kfree() 
operation
to manager those skb data buffers instead of recycling them.
3. if any DATA buffer is available, Instead of calling kmalloc(), assign 
a
skb data buffer directly from ring buffer to the incoming packets.
4. If ring buffer still has space, Instead of calling kfree(), store the 
skb
data buffer into the ring buffer.
5. if the head and tail pointer overlap and head pointer is not empty, 
just
stop accpeting more DATA buffer until some DATA buffer is used for the
incoming packets.

I tested my method on the latest Linux kernel 2.6.13.3, it works with the
normal traffic; However, the Linux kernel crashed under the heavy network
traffic.

Any idea to make this ring bufer work under the heavy network traffic?

Your idea seems interesting, but you forgot to post a link to the code you 
wrote. How do you want us to help you ?

Eric

LeoY a écrit :

Here is the modifications I made:
//skbuff.h
//I added the folloing definitions
#define MAX_POOL_SIZE 4096
unsigned char *skbuff_data_pool[MAX_POOL_SIZE];
int skbPoolHead, skbPoolTail;


//skb_init()
//I add the following codes:
for(i=0;i<MAX_POOL_SIZE;i++)
        skbuff_data_pool[i] = NULL;
skbPoolHead = skbPoolTail = 0;


//alloc_skb()
//I made the following changes
if (skbuff_data_pool[skbPoolHead])
{
  data = skbuff_data_pool[skbPoolHead];
 skbuff_data_pool[skbPoolHead] = NULL;
 if (++skbPoolHead == MAX_POOL_SIZE)
  skbPoolHead = 0;
}
else{ //Original path
 size = SKB_DATA_ALIGN(size);
 data = kmalloc(size + sizeof(struct skb_shared_info), gfp_mask);
 }

//skb_release_data()
//I made the following changes:
 if ((skbPoolHead == skbPoolTail) && (skbuff_data_pool[skbPoolHead] != 
NULL))
 //Original path
  kfree(skb->head);
               else{
  if (skbuff_data_pool[skbPoolTail])
   panic("Tail pointer must be null");

  skbuff_data_pool[skbPoolTail] = skb->head;
                       if (++skbPoolTail == MAX_POOL_SIZE)
   skbPoolTail = 0;
 }

Hum... Lot of problems I think

Are you aware that skb_alloc() / skb_free() can handle data buffers of 
different sizes ?

So if you kmalloc() a small buffer, and store it later in your ring buffer, 
you should not give it back to a caller that need a biger buffer.

If you want your 'ring buffer'to work, I suspect you should ignore the 'size' 
and let it be the max possible size handled in your machine.

Then, dont forget about SMP and IRQs : I dont see in your code how you protect 
against concurrent processors accessing your ring buffers, and how you protect 
against IRQ (since a nic handler can runs on IRQ or softirq context)

kmalloc()/kfree() are SMP safe.

----- Original Message ----- From: "Eric Dumazet" [off-list ref]
To: "LeoY" <redacted>
Cc: <redacted>
Sent: Saturday, October 08, 2005 2:47 PM
Subject: Re: Is this skb recycle buffer helpful to improve Linux network 
stack performance?

quoted

LeoY a écrit :

quoted

Hi,

Motivation: we noticed alloc_skb()/kfree() used lots of clock ticks when
handling heavy network traffic. As Linux kernel always need to call
kmalloc()/kfree() to allocate and deallocate a skb DATA buffer(not 
sk_buff)
for each incoming/outgoing packet, we try to reduce the frequence of 
calling
these memory functions.

I wangt to set up a ring buffer in Linux kernel(skbuff.c) and recycle 
those
skb data buffers. The basic idea is as follows:
1. Create a ring buffer. This ring buffer has a head pointer which 
points to
the virtual address of the data buffer to be reused; It also has a tail
pointer, which can be used to store the virutal address of skb data 
buffer
for those transmitted packets.
2. If the ring buffer is full, just use normal kmalloc()/kfree() 
operation
to manager those skb data buffers instead of recycling them.
3. if any DATA buffer is available, Instead of calling kmalloc(), 
assign a
skb data buffer directly from ring buffer to the incoming packets.
4. If ring buffer still has space, Instead of calling kfree(), store 
the skb
data buffer into the ring buffer.
5. if the head and tail pointer overlap and head pointer is not 
empty, just
stop accpeting more DATA buffer until some DATA buffer is used for the
incoming packets.

I tested my method on the latest Linux kernel 2.6.13.3, it works with 
the
normal traffic; However, the Linux kernel crashed under the heavy 
network
traffic.

Any idea to make this ring bufer work under the heavy network traffic?

Your idea seems interesting, but you forgot to post a link to the code 
you wrote. How do you want us to help you ?

Eric

Hi Eric,

Thanks for your reply.
1. For the packet size, this idea is targeting some specific network card 
driver and we assume all the packet size will not exceed 2KB.
2. Currently only Uni-Processor is considered(Hyper-threading is also 
disabled), I will add spin lock part once it works on the UP.

Leo

----- Original Message ----- 
From: "Eric Dumazet" <redacted>
To: "LeoY" <redacted>
Cc: <redacted>; <redacted>
Sent: Saturday, October 08, 2005 3:09 PM
Subject: Re: Is this skb recycle buffer helpful to improve Linux network 
stack performance?

LeoY a écrit :

quoted

Here is the modifications I made:
//skbuff.h
//I added the folloing definitions
#define MAX_POOL_SIZE 4096
unsigned char *skbuff_data_pool[MAX_POOL_SIZE];
int skbPoolHead, skbPoolTail;


//skb_init()
//I add the following codes:
for(i=0;i<MAX_POOL_SIZE;i++)
        skbuff_data_pool[i] = NULL;
skbPoolHead = skbPoolTail = 0;


//alloc_skb()
//I made the following changes
if (skbuff_data_pool[skbPoolHead])
{
  data = skbuff_data_pool[skbPoolHead];
 skbuff_data_pool[skbPoolHead] = NULL;
 if (++skbPoolHead == MAX_POOL_SIZE)
  skbPoolHead = 0;
}
else{ //Original path
 size = SKB_DATA_ALIGN(size);
 data = kmalloc(size + sizeof(struct skb_shared_info), gfp_mask);
 }

//skb_release_data()
//I made the following changes:
 if ((skbPoolHead == skbPoolTail) && (skbuff_data_pool[skbPoolHead] != 
NULL))
 //Original path
  kfree(skb->head);
               else{
  if (skbuff_data_pool[skbPoolTail])
   panic("Tail pointer must be null");

  skbuff_data_pool[skbPoolTail] = skb->head;
                       if (++skbPoolTail == MAX_POOL_SIZE)
   skbPoolTail = 0;
 }

Hum... Lot of problems I think

Are you aware that skb_alloc() / skb_free() can handle data buffers of 
different sizes ?

So if you kmalloc() a small buffer, and store it later in your ring 
buffer, you should not give it back to a caller that need a biger buffer.

If you want your 'ring buffer'to work, I suspect you should ignore the 
'size' and let it be the max possible size handled in your machine.

Then, dont forget about SMP and IRQs : I dont see in your code how you 
protect against concurrent processors accessing your ring buffers, and how 
you protect against IRQ (since a nic handler can runs on IRQ or softirq 
context)

kmalloc()/kfree() are SMP safe.

quoted

----- Original Message ----- From: "Eric Dumazet" [off-list ref]
To: "LeoY" <redacted>
Cc: <redacted>
Sent: Saturday, October 08, 2005 2:47 PM
Subject: Re: Is this skb recycle buffer helpful to improve Linux network 
stack performance?

quoted

LeoY a écrit :

quoted

Hi,

Motivation: we noticed alloc_skb()/kfree() used lots of clock ticks 
when
handling heavy network traffic. As Linux kernel always need to call
kmalloc()/kfree() to allocate and deallocate a skb DATA buffer(not 
sk_buff)
for each incoming/outgoing packet, we try to reduce the frequence of 
calling
these memory functions.

I wangt to set up a ring buffer in Linux kernel(skbuff.c) and recycle 
those
skb data buffers. The basic idea is as follows:
1. Create a ring buffer. This ring buffer has a head pointer which 
points to
the virtual address of the data buffer to be reused; It also has a tail
pointer, which can be used to store the virutal address of skb data 
buffer
for those transmitted packets.
2. If the ring buffer is full, just use normal kmalloc()/kfree() 
operation
to manager those skb data buffers instead of recycling them.
3. if any DATA buffer is available, Instead of calling kmalloc(), 
assign a
skb data buffer directly from ring buffer to the incoming packets.
4. If ring buffer still has space, Instead of calling kfree(), store 
the skb
data buffer into the ring buffer.
5. if the head and tail pointer overlap and head pointer is not empty, 
just
stop accpeting more DATA buffer until some DATA buffer is used for the
incoming packets.

I tested my method on the latest Linux kernel 2.6.13.3, it works with 
the
normal traffic; However, the Linux kernel crashed under the heavy 
network
traffic.

Any idea to make this ring bufer work under the heavy network traffic?

Your idea seems interesting, but you forgot to post a link to the code 
you wrote. How do you want us to help you ?

Eric

quoted

LeoY a écrit :

Hi Eric,

Thanks for your reply.
1. For the packet size, this idea is targeting some specific network 
card driver and we assume all the packet size will not exceed 2KB.
2. Currently only Uni-Processor is considered(Hyper-threading is also 
disabled), I will add spin lock part once it works on the UP.

1) Even on Uni-Processor, you still need to protect against IRQS

2) The big cost of kmalloc()/kfree() come from the 
local_irq_save(flags)/local_irq_restore(flags)

Once you add them in your code, you will discover you gain nothing compared to 
kmalloc()/kfree() that already use a 'ring buffer' : More over, the slab 
implementation is designed to have separate 'ring buffer' (one for each cpu), 
so it is probably better than a 'central ring buffer'

Eric