LinuxLists.cc - Network compatibility and performance

2006-08-10 15:34:26

2006-08-10 17:29:29

Subject: Re: Network compatibility and performance

On Thu, 10 Aug 2006 11:34:23 -0400
"linux-os \(Dick Johnson\)" <[email protected]> wrote:

>
> Hello,
>
> Network throughput is seriously defective with linux-2.6.16.24
> if the length given to 'write()' is a large number.
>
> Given this code on a connected socket........

What protocol (TCP?) and what Ethernet hardware (does it support TSO)?
Did you set non-blocking?

2006-08-10 18:09:46

by linux-os (Dick Johnson)

[permalink] [raw]

Subject: Re: Network compatibility and performance

On Thu, 10 Aug 2006, Stephen Hemminger wrote:

> On Thu, 10 Aug 2006 11:34:23 -0400
> "linux-os \(Dick Johnson\)" <[email protected]> wrote:
>
>>
>> Hello,
>>
>> Network throughput is seriously defective with linux-2.6.16.24
>> if the length given to 'write()' is a large number.
>>
>> Given this code on a connected socket........
>
> What protocol (TCP?) and what Ethernet hardware (does it support TSO)?
> Did you set non-blocking?

A connected TCP socket. The Ethernet hardware was also
described (Intel using e1000 as shown) It's on PCI-X 133MHz, two
devices on the motherboard, not really relevent because it worked
previously as described. TSO? No ARPA virtual terminals here.
They went away in 1972. The socket was set to non-blocking because the
same socket is used for reading (not at the same time), using poll()
to find when data are supposed to be available. BTW, read() code
used to use poll() to find out when data were available, but if
poll returned POLLIN, sometimes data would NOT be available and
the code would hang <forever>. Therefore a work-around was to set
the socket non-blocking. Under the conditions where poll() would
return POLLIN and a read of a non-blocking socket returned no data,
the errno was 3 (no such process) which seems really strange.
Nevertheless, this has worked in the field for quite some time.

There are no threads or child processes sharing data. Just a
task receiving messages (commands) and then sending data (responses).
Nothing in any user code overlaps.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.16.24 on an i686 machine (5592.62 BogoMips).
New book: http://www.AbominableFirebug.com/
_

****************************************************************
The information transmitted in this message is confidential and may be privileged. Any review, retransmission, dissemination, or other use of this information by persons or entities other than the intended recipient is prohibited. If you are not the intended recipient, please notify Analogic Corporation immediately - by replying to this message or by sending an email to [email protected] - and destroy all copies of this information, including any attachments, without reading or disclosing them.

Thank you.

2006-08-10 18:14:37

by Stephen Hemminger

[permalink] [raw]

Subject: Re: Network compatibility and performance

On Thu, 10 Aug 2006 14:09:34 -0400
"linux-os \(Dick Johnson\)" <[email protected]> wrote:

>
> On Thu, 10 Aug 2006, Stephen Hemminger wrote:
>
> > On Thu, 10 Aug 2006 11:34:23 -0400
> > "linux-os \(Dick Johnson\)" <[email protected]> wrote:
> >
> >>
> >> Hello,
> >>
> >> Network throughput is seriously defective with linux-2.6.16.24
> >> if the length given to 'write()' is a large number.
> >>
> >> Given this code on a connected socket........
> >
> > What protocol (TCP?) and what Ethernet hardware (does it support TSO)?
> > Did you set non-blocking?
>
> A connected TCP socket. The Ethernet hardware was also
> described (Intel using e1000 as shown) It's on PCI-X 133MHz, two
> devices on the motherboard, not really relevent because it worked
> previously as described. TSO?

TSO = TCP segmentation Offload, if you are using e1000 it gets enabled.
Only slightly relevant to this, because it would change the timing.

> They went away in 1972. The socket was set to non-blocking because the
> same socket is used for reading (not at the same time), using poll()
> to find when data are supposed to be available. BTW, read() code
> used to use poll() to find out when data were available, but if
> poll returned POLLIN, sometimes data would NOT be available and
> the code would hang <forever>. Therefore a work-around was to set
> the socket non-blocking. Under the conditions where poll() would
> return POLLIN and a read of a non-blocking socket returned no data,

Basic unix programming, errno only has meaning if system call returns -1.

Basic network programming. If read returns 0 it means other side
has disconnected.

2006-08-10 18:32:37

by linux-os (Dick Johnson)

[permalink] [raw]

Subject: Re: Network compatibility and performance

On Thu, 10 Aug 2006, Stephen Hemminger wrote:

> On Thu, 10 Aug 2006 14:09:34 -0400
> "linux-os \(Dick Johnson\)" <[email protected]> wrote:
>
>>
>> On Thu, 10 Aug 2006, Stephen Hemminger wrote:
>>
>>> On Thu, 10 Aug 2006 11:34:23 -0400
>>> "linux-os \(Dick Johnson\)" <[email protected]> wrote:
>>>
>>>>
>>>> Hello,
>>>>
>>>> Network throughput is seriously defective with linux-2.6.16.24
>>>> if the length given to 'write()' is a large number.
>>>>
>>>> Given this code on a connected socket........
>>>
>>> What protocol (TCP?) and what Ethernet hardware (does it support TSO)?
>>> Did you set non-blocking?
>>
>> A connected TCP socket. The Ethernet hardware was also
>> described (Intel using e1000 as shown) It's on PCI-X 133MHz, two
>> devices on the motherboard, not really relevent because it worked
>> previously as described. TSO?
>
> TSO = TCP segmentation Offload, if you are using e1000 it gets enabled.
> Only slightly relevant to this, because it would change the timing.
>
>> They went away in 1972. The socket was set to non-blocking because the
>> same socket is used for reading (not at the same time), using poll()
>> to find when data are supposed to be available. BTW, read() code
>> used to use poll() to find out when data were available, but if
>> poll returned POLLIN, sometimes data would NOT be available and
>> the code would hang <forever>. Therefore a work-around was to set
>> the socket non-blocking. Under the conditions where poll() would
>> return POLLIN and a read of a non-blocking socket returned no data,
>
> Basic unix programming, errno only has meaning if system call returns -1.
>

True. So what?

> Basic network programming. If read returns 0 it means other side
> has disconnected.
>

So, the other side did not disconnect or shutdown. This is a previously
reported problem that has required work-arounds because it has never
been fixed. The read() problem may not be relevent to the spin-forever
when the kernel doesn't try to send even 1440 bites of the 0x01000000
byte buffer. That's the new problem being reported at this time.

Remember we (me and the others whose code I help review and sometimes
fix) have been doing socket programming since the days BSD first added
them to Unix, and have been comming up with work-arounds for things that
keep changing ever since. We have socket programming running on Interactive
Unix (now that's old), SCO Unix (sorry for the swear word), and SunOS.
I am reporting that something has CHANGED and, even with a work-around
in place, the result is a severe reduction in throughput.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.16.24 on an i686 machine (5592.62 BogoMips).
New book: http://www.AbominableFirebug.com/
_

****************************************************************
The information transmitted in this message is confidential and may be privileged. Any review, retransmission, dissemination, or other use of this information by persons or entities other than the intended recipient is prohibited. If you are not the intended recipient, please notify Analogic Corporation immediately - by replying to this message or by sending an email to [email protected] - and destroy all copies of this information, including any attachments, without reading or disclosing them.

Thank you.

2006-08-12 19:34:13

by Ben Greear

[permalink] [raw]

Subject: Re: Network compatibility and performance

linux-os (Dick Johnson) wrote:
> Hello,
>
> Network throughput is seriously defective with linux-2.6.16.24
> if the length given to 'write()' is a large number.
>
> Given this code on a connected socket........
>
> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
> //
> // Copyright(c) 2005 Analogic Corporation ([email protected])
> //
> // This program may be distributed under the GNU Public License
> // version 2, as published by the Free Software Foundation, Inc.,
> // 59 Temple Place, Suite 330 Boston, MA, 02111.
> //
> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
>
> #include <stdio.h>
> #include <unistd.h>
> #include <stdlib.h>
> #include <stdint.h>
> #include <signal.h>
> #include <string.h>
> #include <stdarg.h>
> #include <errno.h>
> #include <fcntl.h>
> #include <netinet/in.h>
> #include <netinet/tcp.h>
> #include <sys/poll.h>
>
> #define BUF_LEN 0x1000
> #define FAIL -1
>
> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
> //
> // This sends a message that could exceed the size of the network buffers.
> // It returns 0 if everything went okay, and FAIL if not.
> //
> int32_t sender(int32_t fd, void *buf, size_t len)
> {
> int32_t ret_val;
> uint8_t *cp;
> cp = (uint8_t *) buf;
> while(len) {
> if((ret_val = write(fd, cp, MIN(len, BUF_LEN))) == FAIL) {
> if(errno == EAGAIN)
> continue;
> return ret_val;
> }
> len -= ret_val;
> cp += ret_val;
> }
> return 0;
> }
>
> It used to work quite well with:
>
> while(len) {
> if((ret_val = write(fd, cp, len)) == FAIL) {
> return ret_val;
> }
> len -= ret_val;
> cp += ret_val;
> }
>
> The network socket layer would return the amount of bytes
> actually sent and the code would walk its way up through the
> buffer. This was the expected behavior for many years.
>
> Then after about Linux-2.6.8, I needed to do:
>
> while(len) {
> if((ret_val = write(fd, cp, len)) == FAIL) {
> if(errno == EAGAIN)
> continue;
> return ret_val;
> }
> len -= ret_val;
> cp += ret_val;
> }
>
> This was because Linux would claim to run out of resources
> even though there was nothing else running on the system.
>
> Now at Linux-2.6.16.24, the code needed to be further modified
> to:
> while(len) {
> if((ret_val = write(fd, cp, MIN(len, 0x1000))) == FAIL) {
> if(errno == EAGAIN)
> continue;
> return ret_val;
> }
> len -= ret_val;
> cp += ret_val;
> }

In the case where you are getting EAGAIN, this is a busy-spin. You
might want to sleep in a select() or similar call as soon as you get
EAGAIN on this socket..or go off and do other work while the OS clears
out the send queue.

Also, from your description, this code should return 0 on success. It
is returning 'ret_val' instead, which should be > 0.

I have no idea why you need to add the MIN() logic..and that seems like
something that should not be required.

> ... or else it would spin <forever> returning 0 with no errno set.
> In all cases, these problems exist when 'len' is a large value, perhaps
> 0x01000000, much greater than Linux could ever find an available
> buffer for. Linux used to send what it could. Now it will just fail to
> send anything at all, returning 0 if it 'feels' like it doesn't want
> to bother. This is not good. With the hacked code, the data throughput
> is about 100,000 bytes per second on a dedicated link. The previous
> code ran 112,000 bytes per second. Once the 'errno' happens, the
> network stumbles to a measley 12,000 bytes per second. This
> breaks our applications.

Even 112kbps sucks on a decent network. What is the speed of your
network, what protocol are you using, if tcp, what is the latency
of your network?

Ben

2006-08-14 11:30:26

by linux-os (Dick Johnson)

[permalink] [raw]

Subject: Re: Network compatibility and performance

On Sat, 12 Aug 2006, Ben Greear wrote:

> linux-os (Dick Johnson) wrote:
>> Hello,
>>
>> Network throughput is seriously defective with linux-2.6.16.24
>> if the length given to 'write()' is a large number.
>>
>> Given this code on a connected socket........
>>
>> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
>> //
>> // Copyright(c) 2005 Analogic Corporation ([email protected])
>> //
>> // This program may be distributed under the GNU Public License
>> // version 2, as published by the Free Software Foundation, Inc.,
>> // 59 Temple Place, Suite 330 Boston, MA, 02111.
>> //
>> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
>>
>> #include <stdio.h>
>> #include <unistd.h>
>> #include <stdlib.h>
>> #include <stdint.h>
>> #include <signal.h>
>> #include <string.h>
>> #include <stdarg.h>
>> #include <errno.h>
>> #include <fcntl.h>
>> #include <netinet/in.h>
>> #include <netinet/tcp.h>
>> #include <sys/poll.h>
>>
>> #define BUF_LEN 0x1000
>> #define FAIL -1
>>
>> //-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
>> //
>> // This sends a message that could exceed the size of the network buffers.
>> // It returns 0 if everything went okay, and FAIL if not.
>> //
>> int32_t sender(int32_t fd, void *buf, size_t len)
>> {
>> int32_t ret_val;
>> uint8_t *cp;
>> cp = (uint8_t *) buf;
>> while(len) {
>> if((ret_val = write(fd, cp, MIN(len, BUF_LEN))) == FAIL) {
>> if(errno == EAGAIN)
>> continue;
>> return ret_val;
>> }
>> len -= ret_val;
>> cp += ret_val;
>> }
>> return 0;
>> }
>>
>> It used to work quite well with:
>>
>> while(len) {
>> if((ret_val = write(fd, cp, len)) == FAIL) {
>> return ret_val;
>> }
>> len -= ret_val;
>> cp += ret_val;
>> }
>>
>> The network socket layer would return the amount of bytes
>> actually sent and the code would walk its way up through the
>> buffer. This was the expected behavior for many years.
>>
>> Then after about Linux-2.6.8, I needed to do:
>>
>> while(len) {
>> if((ret_val = write(fd, cp, len)) == FAIL) {
>> if(errno == EAGAIN)
>> continue;
>> return ret_val;
>> }
>> len -= ret_val;
>> cp += ret_val;
>> }
>>
>> This was because Linux would claim to run out of resources
>> even though there was nothing else running on the system.
>>
>> Now at Linux-2.6.16.24, the code needed to be further modified
>> to:
>> while(len) {
>> if((ret_val = write(fd, cp, MIN(len, 0x1000))) == FAIL) {
>> if(errno == EAGAIN)
>> continue;
>> return ret_val;
>> }
>> len -= ret_val;
>> cp += ret_val;
>> }
>
> In the case where you are getting EAGAIN, this is a busy-spin. You
> might want to sleep in a select() or similar call as soon as you get
> EAGAIN on this socket..or go off and do other work while the OS clears
> out the send queue.
>
> Also, from your description, this code should return 0 on success. It
> is returning 'ret_val' instead, which should be > 0.

No it will return FAIL (-1) or an error and 0 (the bottom of the procedure)
if the whole things went. It is mandatory that the whole thing goes
so this procedure should handle any intermediate actions.

Upon your advice, I may try to add select() although, on a write it
seems to be putting in user-space something that used to be handled
quite well in the kernel. I don't think the user should really care
about the kernel internals, whether or not the kernel happens to have
a buffer available.

>
> I have no idea why you need to add the MIN() logic..and that seems like
> something that should not be required.
>

It seems that some code 'thinks' that a large buffer of data is
an error and won't even try to send some anymore.

>> ... or else it would spin <forever> returning 0 with no errno set.
>> In all cases, these problems exist when 'len' is a large value, perhaps
>> 0x01000000, much greater than Linux could ever find an available
>> buffer for. Linux used to send what it could. Now it will just fail to
>> send anything at all, returning 0 if it 'feels' like it doesn't want
>> to bother. This is not good. With the hacked code, the data throughput
>> is about 100,000 bytes per second on a dedicated link. The previous
>> code ran 112,000 bytes per second. Once the 'errno' happens, the
>> network stumbles to a measley 12,000 bytes per second. This
>> breaks our applications.
>
> Even 112kbps sucks on a decent network. What is the speed of your
> network, what protocol are you using, if tcp, what is the latency
> of your network?
>

The network is a single wire about 8 feet long, connecting Intel gigibit
links on two identical computers (crossover cable). This link is TCP.
For high-speed data, I use UDP and I get a higher throughput because
there is no handshake. Thew latency is the latency of Linux. BTW, it's
only a gigaBIT link, you can divide that by 8 for gigabytes. I don't
know the actual bit-rate on the wires, if we assume 1GHz, the byte-rate
is only 125,000 bytes per second. Being able to use 89.6 percent of
that isn't bad at all.

> Ben
>
>
>

Cheers,
Dick Johnson
Penguin : Linux version 2.6.16.24 on an i686 machine (5592.62 BogoMips).
New book: http://www.AbominableFirebug.com/
_

****************************************************************
The information transmitted in this message is confidential and may be privileged. Any review, retransmission, dissemination, or other use of this information by persons or entities other than the intended recipient is prohibited. If you are not the intended recipient, please notify Analogic Corporation immediately - by replying to this message or by sending an email to [email protected] - and destroy all copies of this information, including any attachments, without reading or disclosing them.

Thank you.

2006-08-14 21:26:31

by Ben Greear

[permalink] [raw]

Subject: Re: Network compatibility and performance

linux-os (Dick Johnson) wrote:

> No it will return FAIL (-1) or an error and 0 (the bottom of the procedure)
> if the whole things went. It is mandatory that the whole thing goes
> so this procedure should handle any intermediate actions.

I see..I missed that part.

> Upon your advice, I may try to add select() although, on a write it
> seems to be putting in user-space something that used to be handled
> quite well in the kernel. I don't think the user should really care
> about the kernel internals, whether or not the kernel happens to have
> a buffer available.

Since you put it in non-blocking mode, you need the select() to throttle
unless you want to busy spin. Whether you should have to actually put
in in non-blocking mode or not is a different question.

>>I have no idea why you need to add the MIN() logic..and that seems like
>>something that should not be required.
>>
>
> It seems that some code 'thinks' that a large buffer of data is
> an error and won't even try to send some anymore.

I have seen a problem where I can repeatedly hang a TCP connection
when running at high speed. The tx queue is full or mostly full, and
on the wire I only see 200kpps of duplicate acks. Can't reproduce it
with anything other than my big complicated proprietary app, so it
remains unfixed.

I am not sure if this is related to what you see or not..but could you
check to see if there is lots of acks on the wire when this hang happens?

>>Even 112kbps sucks on a decent network. What is the speed of your
>>network, what protocol are you using, if tcp, what is the latency
>>of your network?
>>
>
>
> The network is a single wire about 8 feet long, connecting Intel gigibit
> links on two identical computers (crossover cable). This link is TCP.
> For high-speed data, I use UDP and I get a higher throughput because
> there is no handshake. Thew latency is the latency of Linux. BTW, it's
> only a gigaBIT link, you can divide that by 8 for gigabytes. I don't
> know the actual bit-rate on the wires, if we assume 1GHz, the byte-rate
> is only 125,000 bytes per second. Being able to use 89.6 percent of
> that isn't bad at all.

You must be meaning to add a few more zeros to that number. If you
are getting ~125,000,000 Bytes per second then you are doing OK.

Ben

2006-08-15 11:34:21

by linux-os (Dick Johnson)

[permalink] [raw]

Subject: Re: Network compatibility and performance

On Mon, 14 Aug 2006, Ben Greear wrote:

> linux-os (Dick Johnson) wrote:
>
>> No it will return FAIL (-1) or an error and 0 (the bottom of the procedure)
>> if the whole things went. It is mandatory that the whole thing goes
>> so this procedure should handle any intermediate actions.
>
> I see..I missed that part.
>
>> Upon your advice, I may try to add select() although, on a write it
>> seems to be putting in user-space something that used to be handled
>> quite well in the kernel. I don't think the user should really care
>> about the kernel internals, whether or not the kernel happens to have
>> a buffer available.
>
> Since you put it in non-blocking mode, you need the select() to throttle
> unless you want to busy spin. Whether you should have to actually put
> in in non-blocking mode or not is a different question.
>
>>> I have no idea why you need to add the MIN() logic..and that seems like
>>> something that should not be required.
>>>
>>
>> It seems that some code 'thinks' that a large buffer of data is
>> an error and won't even try to send some anymore.
>
> I have seen a problem where I can repeatedly hang a TCP connection
> when running at high speed. The tx queue is full or mostly full, and
> on the wire I only see 200kpps of duplicate acks. Can't reproduce it
> with anything other than my big complicated proprietary app, so it
> remains unfixed.
>
> I am not sure if this is related to what you see or not..but could you
> check to see if there is lots of acks on the wire when this hang happens?

I will check to see what it's doing on the wire and get back.

>
>>> Even 112kbps sucks on a decent network. What is the speed of your
>>> network, what protocol are you using, if tcp, what is the latency
>>> of your network?
>>>
>>
>>
>> The network is a single wire about 8 feet long, connecting Intel gigibit
>> links on two identical computers (crossover cable). This link is TCP.
>> For high-speed data, I use UDP and I get a higher throughput because
>> there is no handshake. Thew latency is the latency of Linux. BTW, it's
>> only a gigaBIT link, you can divide that by 8 for gigabytes. I don't
>> know the actual bit-rate on the wires, if we assume 1GHz, the byte-rate
>> is only 125,000 bytes per second. Being able to use 89.6 percent of
>> that isn't bad at all.
>
> You must be meaning to add a few more zeros to that number. If you
> are getting ~125,000,000 Bytes per second then you are doing OK.
>

ACK that!

> Ben
>
>

Cheers,
Dick Johnson
Penguin : Linux version 2.6.16.24 on an i686 machine (5592.62 BogoMips).
New book: http://www.AbominableFirebug.com/
_

****************************************************************
The information transmitted in this message is confidential and may be privileged. Any review, retransmission, dissemination, or other use of this information by persons or entities other than the intended recipient is prohibited. If you are not the intended recipient, please notify Analogic Corporation immediately - by replying to this message or by sending an email to [email protected] - and destroy all copies of this information, including any attachments, without reading or disclosing them.

Thank you.

Subject: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance

Subject: Re: Network compatibility and performance