RE: 923Mbits/s across the ocean
Also as the OS's are shipped they come with small default maximum window sizes (I think Linux is typically 64KB and Solaris is 8K), and so one has to get the sysadmin with root privs to change this. -----Original Message----- From: Iljitsch van Beijnum [mailto:iljitsch@muada.com] Sent: Sunday, March 09, 2003 5:25 AM To: Joe St Sauver Cc: nanog@nanog.org Subject: Re: 923Mbits/s across the ocean On Sat, 8 Mar 2003, Joe St Sauver wrote:
you will see that for bulk TCP flows, the median throughput is still only 2.3Mbps. 95th%-ile is only ~9Mbps. That's really not all that great, throughput wise, IMHO.
Strange. Why is that? RFC 1323 is widely implemented, although not widely enabled (and for good reason: the timestamp option kills header compression so it's bad for lower-bandwidth connections). My guess is that the OS can't afford to throw around MB+ size buffers for every TCP session so the default buffers (which limit the windows that can be used) are relatively small and application programmers don't override the default.
On Sun, Mar 09, 2003 at 08:29:16AM -0800, Cottrell, Les wrote:
Strange. Why is that? RFC 1323 is widely implemented, although not widely enabled (and for good reason: the timestamp option kills header compression so it's bad for lower-bandwidth connections). My guess is that the OS can't afford to throw around MB+ size buffers for every TCP session so the default buffers (which limit the windows that can be used) are relatively small and application programmers don't override the default.
Also as the OS's are shipped they come with small default maximum window sizes (I think Linux is typically 64KB and Solaris is 8K), and so one has to get the sysadmin with root privs to change this.
This is related to how the kernel/user model works in relation to TCP. TCP itself happens in the kernel, but the data comes from userland through the socket interface, so there is a "socket buffer" in the kernel which holds data coming from and going to the application. TCP cannot release data from it's buffer until it has been acknowledged by the other side, incase it needs to retransmit. This means TCP performance is limited by the smaller of either the congestion window (determined by measuring conditions along the path), or the send/recv window (determined by local system resources). However, you can't just blindly turn up your socket buffers to large values and expect good results. On the send size, the application transmitting is guaranteed to utilize the buffers immediately (ever seen a huge jump in speed at the beginning of a transfer, this is the local buffer being filled, and the application has no way to know if this data is going out to the wire, or just to the kernel). Then the network must drain the packets onto the wire, sometimes very slowly (think about a dialup user downloading from your GigE server). Setting the socket buffers too high can potentially result in an incredible waste of resources, and can severely limit the number of simultaneous connections your server can support. This is precisely why OS's cannot ship with huge default values, because what may be appropriate for your one-user GigE connected box might not be appropriate for someone else's 100BASE-TX web server (and guess which setup has more users :P). On the receive size, the socket buffers must be large enough to accommodate all the data received between application read()'s, as well as making sure they have enough available space to hold future data in the event of a "gap" due to loss and the need for retransmission. However, if the application fails to read() the data from the socket buffer, it will sit there forever. Large socket buffers also opens the server up to malicious attack causing non-swapable kernel memory to consume all available resources, either locally (by someone dumping data over lots of connections, or running an application which intentionally fails to read data from the socket buffer), or remotely (think someone opening a bunch of rate limited connections from your "high speed server"). It can even be unintentional, but just as bad (think a million confused dialup users accidentally clicking on your high speed video stream). Some of this can be worked around by implementing what is called auto-tuning socket buffers. In this case, the kernel would limit the amount of data allowed into the buffer, by looking at the tcp session's observed congestion window. This allows you to define large send buffers without applications connected to slow receivers sucking up unnecessary resourced. PSC has had example implementations for quite a while, and recently FreeBSD even added this (sysctl net.inet.tcp.inflight_enable=1 as of 4.7). Unfortunately, there isn't much you can do to prevent malicious receive-side buffer attacks, short of limiting the overall max buffer (FreeBSD implements this as an rlimit "sbsize"). Of course, you need a few other things before you can start getting into end to end gigabit speeds. If you're transfering a file, you probably don't want to be reading it from disk via the kernel just to send it back to the kernel again for transmission, so various things like sendfile() and zero copy implementations help get you the performance you need locally. Jumbo frames help too, but their real benefit is not the simplistic "hey look theres 1/3rd the number of frames/sec" view that many people see. The good stuff comes from techniques like page flipping, where the NIC DMA's data into a memory page which can be flipped through the system straight to the application, without copying it throughout. Some day TCP may just be implemented on the NIC itself, with ALL work offloaded, and the system doing nothing but receiving nice page-sized chunks of data at high rates of speed. IMHO the 1500 byte MTU of ethernet will still continue to prevent good end to end performance like this for a long time to come. But alas, I digress... -- Richard A Steenbergen <ras@e-gerbil.net> http://www.e-gerbil.net/ras GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)
On Sun, 9 Mar 2003, Richard A Steenbergen wrote:
On the send size, the application transmitting is guaranteed to utilize the buffers immediately (ever seen a huge jump in speed at the beginning of a transfer, this is the local buffer being filled, and the application has no way to know if this data is going out to the wire, or just to the kernel). Then the network must drain the packets onto the wire, sometimes very slowly (think about a dialup user downloading from your GigE server).
Actually this is often way too fast as the congestion window doubles with each ACK. This means that with a large buffer = large window and a bottleneck somewhere along the way, you are almost guaranteed to have some serious congestion in the early stages of the session and lower levels of congestion periodially later on whenever TCP tries to figure out how large the congestion window can get without losing packets. This is the part about TCP that I've never understood: why does it send large numbers of packets back-to-back? This is almost never a good idea.
On the receive size, the socket buffers must be large enough to accommodate all the data received between application read()'s,
That's not true. It's perfectly acceptable for TCP to stall when the receiving application fails to read the data fast enough. (TCP then simply announces a window of 0 to the other side so the communication effectively stops until the application reads some data and a >0 window is announced.) If not, the kernel would be required to buffer unlimited amounts of data in the event an application fails to read it from the buffer for some time (which is a very common situation).
locally. Jumbo frames help too, but their real benefit is not the simplistic "hey look theres 1/3rd the number of frames/sec" view that many people see. The good stuff comes from techniques like page flipping, where the NIC DMA's data into a memory page which can be flipped through the system straight to the application, without copying it throughout. Some day TCP may just be implemented on the NIC itself, with ALL work offloaded, and the system doing nothing but receiving nice page-sized chunks of data at high rates of speed.
Hm, I don't see this happening to a usable degree as TCP has no concept of records. You really want to use fixed size chunks of information here rather than pretending everything's a stream.
IMHO the 1500 byte MTU of ethernet will still continue to prevent good end to end performance like this for a long time to come. But alas, I digress...
Don't we all? I'm afraid you're right. Anyone up for modifying IPv6 ND to support a per-neighbor MTU? This should make backward-compatible adoption of jumboframes a possibility. (Maybe retrofit ND into v4 while we're at it.) Iljitsch van Beijnum
On Tue, Mar 11, 2003 at 12:41:15AM +0100, Iljitsch van Beijnum wrote:
On the receive size, the socket buffers must be large enough to accommodate all the data received between application read()'s,
That's not true. It's perfectly acceptable for TCP to stall when the receiving application fails to read the data fast enough. (TCP then simply announces a window of 0 to the other side so the communication effectively stops until the application reads some data and a >0 window is announced.) If not, the kernel would be required to buffer unlimited amounts of data in the event an application fails to read it from the buffer for some time (which is a very common situation).
Ok, I think I was unclear. You don't NEED to have buffers large enough to accommodate all that data received between application read()'s, unless you are trying to achieve maximum performance. I thought that was the general framework we were all working under. :)
locally. Jumbo frames help too, but their real benefit is not the simplistic "hey look theres 1/3rd the number of frames/sec" view that many people see. The good stuff comes from techniques like page flipping, where the NIC DMA's data into a memory page which can be flipped through the system straight to the application, without copying it throughout. Some day TCP may just be implemented on the NIC itself, with ALL work offloaded, and the system doing nothing but receiving nice page-sized chunks of data at high rates of speed.
Hm, I don't see this happening to a usable degree as TCP has no concept of records. You really want to use fixed size chunks of information here rather than pretending everything's a stream.
We're talking optimizations for high performance transfers... It can't always be a stream.
IMHO the 1500 byte MTU of ethernet will still continue to prevent good end to end performance like this for a long time to come. But alas, I digress...
Don't we all? I'm afraid you're right. Anyone up for modifying IPv6 ND to support a per-neighbor MTU? This should make backward-compatible adoption of jumboframes a possibility. (Maybe retrofit ND into v4 while we're at it.)
Not necessarily sure thats the right thing to do, but SOMETHIG has got to be better than what passes for path mtu discovery now. :) -- Richard A Steenbergen <ras@e-gerbil.net> http://www.e-gerbil.net/ras GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)
On Mon, 10 Mar 2003, Richard A Steenbergen wrote:
On the receive size, the socket buffers must be large enough to accommodate all the data received between application read()'s,
That's not true. It's perfectly acceptable for TCP to stall when the receiving application fails to read the data fast enough.
Ok, I think I was unclear. You don't NEED to have buffers large enough to accommodate all that data received between application read()'s, unless you are trying to achieve maximum performance. I thought that was the general framework we were all working under. :)
You got me there. :-) It seemed that you were talking about more general requirements at this point, though with the upper and lower limits for kernel buffer space and all.
Hm, I don't see this happening to a usable degree as TCP has no concept of records. You really want to use fixed size chunks of information here rather than pretending everything's a stream.
We're talking optimizations for high performance transfers... It can't always be a stream.
Right. But TCP is a stream protocol. This has many advantages, nearly all of which are irrelevant for high volume high bandwidth bulk data transfer. I can imagine a system that only works in one direction and where the data is split into fixed size records (which would ideally fit into a single packet) where each record is acknowledged independently (but certainly not for each individual packet). I would also want to take advantage of traffic classification mechanisms: first the data is flooded at the maximum speed at the lowest possible traffic class. Everything that doesn't make it to the other end is then resent slower with a higher traffic class. If the network supports priority queuing then this would effectively sponge up all free bandwidth without impacting regular interactive traffic. If after a few retries some data still didn't make it: simply skip this for now (but keep a record of the missing bits) and keep going. Many applications can live with some lost data and for others it's probably more efficient to keep running at high speed and repair the gaps afterwards.
IMHO the 1500 byte MTU of ethernet will still continue to prevent good end to end performance like this for a long time to come. But alas, I digress...
Don't we all? I'm afraid you're right. Anyone up for modifying IPv6 ND to support a per-neighbor MTU? This should make backward-compatible adoption of jumboframes a possibility. (Maybe retrofit ND into v4 while we're at it.)
Not necessarily sure thats the right thing to do, but SOMETHIG has got to be better than what passes for path mtu discovery now. :)
We can't replace path MTU discovery (but hopefully people will start to realize ICMP messages were invented for another reason than job security for firewalls). But what we need is a way for 10/100 Mbps 1500 byte hosts to live with 1000 Mbps 9000 byte hosts on the same subnet. I thought IPv6 neighbor discovery supported this because ND can communicate the MTU between hosts on the same subnet, but unfortunately this is a subnet-wide MTU and not a per-host MTU, which is what we really need. Iljitsch
[just discovered in my unsent messages queue from offline composition, probably not timely, but...] Iljitsch van Beijnum wrote:
We can't replace path MTU discovery (but hopefully people will start to realize ICMP messages were invented for another reason than job security for firewalls). But what we need is a way for 10/100 Mbps 1500 byte hosts to live with 1000 Mbps 9000 byte hosts on the same subnet. I thought IPv6 neighbor discovery supported this because ND can communicate the MTU between hosts on the same subnet, but unfortunately this is a subnet-wide MTU and not a per-host MTU, which is what we really need.
A decade ago, when I designed SIPP Neighbor Discovery, it saved per destination "maximum unfragmented datagram size" in the route cache, and each I-Am-Here message Heard specified Maximum Receive Unit (MRU) per host. Thus, once upon a time, IPv6 had what you need. Unfortunately, the IPv6 group stripped out such innovative features. I stopped paying attention after the new editor stated something like "it worked for ethernet, we really don't need any more than that." Well, we used IPv4 from '83, and designed SIPP (cum IPv6) in '93. IPv6 is a failure -- maybe it's time for this decade's design? Or maybe even some of the features some of us thought we needed a decade ago? -- William Allen Simpson Key fingerprint = 17 40 5E 67 15 6F 31 26 DD 0D B9 9B 6A 15 2C 32
Thus spake "Iljitsch van Beijnum" <iljitsch@muada.com>
This is the part about TCP that I've never understood: why does it send large numbers of packets back-to-back? This is almost never a good idea.
Because until you congest the network to the point of dropping packets, a host has no idea how much bw is actually available. Exponential rate growith finds this value very quickly.
Hm, I don't see this happening to a usable degree as TCP has no concept of records. You really want to use fixed size chunks of information here rather than pretending everything's a stream.
A record-oriented, reliable transport would make many protocols much easier to implement. Too bad SCTP hasn't seen wider use. S Stephen Sprunk "God does not play dice." --Albert Einstein CCIE #3723 "God is an inveterate gambler, and He throws the K5SSS dice at every possible opportunity." --Stephen Hawking
participants (5)
-
Cottrell, Les -
Iljitsch van Beijnum -
Richard A Steenbergen -
Stephen Sprunk -
William Allen Simpson