search for: 800mbps

Now that we have a host timer based tx wakeup it waits for 64 packets or timeout before processing them. This might cause the guest to run out of tx buffers while the host holds them up. This is a temporal solution to quickly bring back performance to 800mbps. But a better fix will soon be sent (its not the only problem). Signed-off-by: Dor Laor <dor.laor@qumranet.com> --- qemu/hw/virtio-net.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/qemu/hw/virtio-net.c b/qemu/hw/virtio-net.c index 777fe2c..3d07b65 100644 --- a/...

Now that we have a host timer based tx wakeup it waits for 64 packets or timeout before processing them. This might cause the guest to run out of tx buffers while the host holds them up. This is a temporal solution to quickly bring back performance to 800mbps. But a better fix will soon be sent (its not the only problem). Signed-off-by: Dor Laor <dor.laor@qumranet.com> --- qemu/hw/virtio-net.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/qemu/hw/virtio-net.c b/qemu/hw/virtio-net.c index 777fe2c..3d07b65 100644 --- a/...

Dave Taht, on Wed 02 Dec 2015 14:41:35 +0100, wrote: > I guess my meta point is driven by my headaches. Getting per packet > processing to scale up past 100Mbit is hard without offloads even on > embedded hardware considered "high end". In my tests I was getting 800Mbps with common laptop Gb ethernet devices. > >> > At least for now we could commit the recvmmsg part? > >> > >> Not my call. It is a linux only thing, so far as I know. > > > > ATM yes. I wouldn't be surprised that other OSes adopt it, though. > >...

...chset updates kvm repository with Anthony's virtio implementation along with rx performance improvements and guest reset handling. The original code was sent to qemu devel list 2 weeks ago. It contains support for network & block devices. Using the performance improvement I was able to do 800Mbps for rx/tx. The performance improvement are not implemented in a qemu standard code but the intention is to first let the community have decent performance with a merged virtio and afterwards do the polishing (there is dma support on the way for qemu). Enjoy [with the courtesy of Anthony & Rus...

...chset updates kvm repository with Anthony's virtio implementation along with rx performance improvements and guest reset handling. The original code was sent to qemu devel list 2 weeks ago. It contains support for network & block devices. Using the performance improvement I was able to do 800Mbps for rx/tx. The performance improvement are not implemented in a qemu standard code but the intention is to first let the community have decent performance with a merged virtio and afterwards do the polishing (there is dma support on the way for qemu). Enjoy [with the courtesy of Anthony & Rus...

10.12.2015 09:20, Michael Tokarev wrote: > 10.12.2015 03:35, Samuel Thibault wrote: > [] > > I suggest reducing ifdeffery in handle_incoming_vpn_data(), especially > the error checking code. > > The function isn't that large now, it might be much better to have > two different implementations. Like this (untested, patch attached): > > void

Hi, we are currently considering different options for rolling out a large scale IP PBX to handle around 3,000 + concurrent calls. Can this be done with Asterisk? Has it been done before? I really would like an input on this. Thanks! -------------- next part -------------- An HTML attachment was scrubbed... URL:

Hi Philipp Thank you for chiming in. The only reason i use the -kh fork is because it seemed to be more recent. I had this performance issue with the regular version and tried the fork. I realize that the problem can be with the TCP stack parameters. I had no problem getting about 800Mbps between these machines when using iperf. Certainly, the workload completely different, but at least i know that the TCP stack is somewhat operational. Do you happen to know specifically what’s broken? I don’t have access to a physical data center this is why i would like to use either EC2 or Azur...

...kets one at a time and also to copy them to a temporal buffer. This patch prevents qemu handlers from reading the tap and instead it selects the tap descriptors for virtio devices. This eliminates copies and also batch guest notifications (interrupts). Using this patch the rx performance reaches 800Mbps. The patch does not follow qemu's api since the intention is first to have a better io in kvm and then to polish it correctly. Signed-off-by: Dor Laor <dor.laor@qumranet.com> --- qemu/hw/pc.h | 2 +- qemu/hw/virtio-net.c | 114 +++++++++++++++++++++++++++++++++++------------...

...kets one at a time and also to copy them to a temporal buffer. This patch prevents qemu handlers from reading the tap and instead it selects the tap descriptors for virtio devices. This eliminates copies and also batch guest notifications (interrupts). Using this patch the rx performance reaches 800Mbps. The patch does not follow qemu's api since the intention is first to have a better io in kvm and then to polish it correctly. Signed-off-by: Dor Laor <dor.laor@qumranet.com> --- qemu/hw/pc.h | 2 +- qemu/hw/virtio-net.c | 114 +++++++++++++++++++++++++++++++++++------------...

...nts in bulk data throughput performance are achieved. In other words, transfers over the internet are a lot faster with this patch. Increase in performance of more than an order of magnitude are pretty common. If you make use of the now included NONE cipher I've hit data rates of more than 800Mbps between Pittsburgh and Chicago and 650Mbps between Pittsburgh and San Diego. The NONE cipher is only used during the bulk data transfer and *not* during authentication. This version of the patch makes the NONE cipher a server side configuration option so you don't have to enable it if you d...

...thus compiled the same as before), and makes the code index into the > arrays. You may want to use interdiff -w /dev/null patch to better see > what changes the patch makes. > > With this patch, I saw the non-ciphered bandwidth achieved over direct > ethernet improve from 680Mbps to 800Mbps (or conversely, reduce the CPU > usage for the same bandwidth). That's great! It would be good though to split handle_incoming_vpn_data() into a function that does the actual recvfrom/mmsg() and one that processes each individual packet, to reduce the level of indentation and make the funct...

Michael Tokarev, on Thu 10 Dec 2015 09:34:45 +0300, wrote: > it is highly unlikely we'll have 256 packets in queue to process. I did get a bit more than a hundred packets in the queue at 800Mbps. But we can reduce to 64, yes. I wouldn't recommend using a static buffer, since we'd want to go threaded at some point. Allocating an array on the stack is very cheap anyway. Samuel

...s kvm repository with Rusty's/Anthony's virtio implementation along with already sent tx performance bug fix and new module reload fixes. The first 9 patches are Rusty's and Anthony's work and are pending in Rusty's tree. This code together with the userspace patches able to do 800Mbps for network rx/tx. Enjoy [with the courtesy of Anthony & Rusty], Dor

...s kvm repository with Rusty's/Anthony's virtio implementation along with already sent tx performance bug fix and new module reload fixes. The first 9 patches are Rusty's and Anthony's work and are pending in Rusty's tree. This code together with the userspace patches able to do 800Mbps for network rx/tx. Enjoy [with the courtesy of Anthony & Rusty], Dor

...sg is not available, and thus compiled the same as before), and makes the code index into the arrays. You may want to use interdiff -w /dev/null patch to better see what changes the patch makes. With this patch, I saw the non-ciphered bandwidth achieved over direct ethernet improve from 680Mbps to 800Mbps (or conversely, reduce the CPU usage for the same bandwidth). More is yet to come: I'll have a look at extending the tun/tap interface to send/receive several packets at a time, and then also using sendmmsg will again improve performance. Samuel --- configure.ac.original 2015-10-02 17:06:31....

On Thu, Dec 10, 2015 at 10:15:19AM +0100, Samuel Thibault wrote: > I did get a bit more than a hundred packets in the queue at 800Mbps. But > we can reduce to 64, yes. I wouldn't recommend using a static buffer, > since we'd want to go threaded at some point. Allocating an array on the > stack is very cheap anyway. I assumed that one would only get more than 1 packet at a time under heavy load, but apparently tr...

Dave Taht, on Wed 02 Dec 2015 14:13:27 +0100, wrote: > More recently Tom Herbert was working on udp encapsulation methods in > the kernel "foo over udp" > > https://www.netdev01.org/docs/herbert-UDP-Encapsulation-Linux.pdf > > https://lwn.net/Articles/614348/ > > which preserve things important at high rates like GRO/GSO. Yes, FOU will probably get the highest

...thus compiled the same as before), and makes the code index into the > arrays. You may want to use interdiff -w /dev/null patch to better see > what changes the patch makes. > > With this patch, I saw the non-ciphered bandwidth achieved over direct > ethernet improve from 680Mbps to 800Mbps (or conversely, reduce the CPU > usage for the same bandwidth). > > More is yet to come: I'll have a look at extending the tun/tap interface > to send/receive several packets at a time, and then also using sendmmsg > will again improve performance. > > Samuel > > ---...