Questions for future enhancement about mtcp HOT 13 CLOSED

Hi Santos

For the questions,
I do not have the exact pps number, but we can guess from transactions/s.

For payloads smaller than 1460B, there will be packets like below.
Client Server
------------- SYN ------------->
<------- SYN/ACK -----------
------------- ACK ------------->
----------- Request ---------->
<------------- ACK -------------
<--------- Response ---------
<------------- FIN -------------
---------- FIN/ACK --------->
<------------- ACK -------------
So 1 transaction takes 5 TX packets, 4 RX packets at max.
Guessing from the Figure 7 in our paper,
since it reaches around 1.5 million transactions/s, we get 7.5 million pps
for TX and 6 million pps for RX at the server side.

About the remaining question, we don't have plan to do them for right now.

Thanks,
EunYoung

2014-05-06 21:22 GMT+09:00 nginxsantos [email protected]:

Hi,
Thanks for putting this great module. The userspace TCP is the hot cake in
the market. I have been looking for one open source for sometime. I have
few questions -

Can you please let me know what PPS you are able to achieve with this.
2.

Do you plan to make it work for multiple processes similar to the
NGINX architecture.
3.

Do you plan to integrate this with netmap.

Thanks..Santos

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1
.

from mtcp.

Thanks. Can you please confirm how many CPUs/ cores were used to get the 1.5 million transactions/s?

from mtcp.

I used 1 CPU which has 8 cores (specifically Intel Xeon E5-2690)
This link will be helpful. It describes our evaluation environment.
http://shader.kaist.edu/mtcp/

EunYoung

2014-05-07 15:35 GMT+09:00 nginxsantos [email protected]:

Thanks. Can you please confirm how many CPUs/ cores were used to get the
1.5 million transactions/s?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42393992
.

from mtcp.

Hi,
Have few more questions -

1. Does this only work with the systems having 10G ethernet cards ?
2. Does this only work for Intel Architecture?

Why I am asking is if I am using this on a Amazon EC2 instance, will that work ?

Regards, Santos

from mtcp.

Sorry I don't think so.

1. As you mentioned this system requires NICs with Intel 82599 chipset.
   In addition, I don't think it will work well on the virtualized environment
   as well.
   The io_engine driver that we use for packet I/O requires them.
2. Any CPUs implementing amd64 will be okay.

Thanks,
EunYoung

2014-05-07 15:45 GMT+09:00 nginxsantos [email protected]:

Hi,
Have few more questions -

1. Does this only work with the systems having 10G ethernet cards ?
2. Does this only work for Intel Architecture?

Why I am asking is if I am using this on a Amazon EC2 instance, will that
work ?

Regards, Santos

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42394585
.

from mtcp.

Thank you. I am still a bit confused.

1. Are you saying that the current support which you have is only for "Intel 82599-based 10 GbE cards with Linux 2.6 kernel". Are you saying that I can't use this with systems having other chipsets, does it work with 1G NIC? Can I compile your io_engine for other linux kernel versions?
2. Are you also saying that you solution does not work with VMs.

Can you please clarify.

Many thanks.
Santos

from mtcp.

1. Yes, it only works with "Intel 82599-based 10 GbE cards with Linux 2.6
   kernel" and io_engine does not compile on the kernels > 3.0
2. Yes, it does not work with VMs.

We know this becomes big limitation and thinking of porting it to DPDK
framework as well.

EunYoung

2014-05-07 16:12 GMT+09:00 nginxsantos [email protected]:

Thank you. I am still a bit confused.

Are you saying that the current support which you have is only for
"Intel 82599-based 10 GbE cards with Linux 2.6 kernel". Are you saying that
I can't use this with systems having other chipsets, does it work with 1G
NIC? Can I compile your io_engine for other linux kernel versions?
2.

Are you also saying that you solution does not work with VMs.

Can you please clarify.

Many thanks.
Santos

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42396236
.

from mtcp.

Thank you.

Just a suggestion -
May be you can consider integrating your stack with netmap. I have tested the performance of "netmap" on a VM with single core, it comes around 700 K PPS. Also, the netmap support is good.

Please let me know if you are proceeding towards a generic approach.

from mtcp.

Sure I will.
Thank you for the suggestion.

EunYoung

2014-05-07 17:39 GMT+09:00 nginxsantos [email protected]:

Thank you.

Just a suggestion -
May be you can consider integrating your stack with netmap. I have tested
the performance of "netmap" on a VM with single core, it comes around 700 K
PPS. Also, the netmap support is good.

Please let me know if you are proceeding towards a generic approach.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42402381
.

from mtcp.

I am also interested for this..an opensource USTCP which can run in VM.

from mtcp.

Could you give me a link for it?
I could not find a reference to it.
Thanks

2014-05-07 17:56 GMT+09:00 petergsnm [email protected]:

I am also interested for this..an opensource USTCP which can run in VM.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42403860
.

from mtcp.

Try this:

git clone https://code.google.com/p/netmap/

You can run without any drivers (the driver improves the performance). It has some examples. You can just see how much PPS can the netmap able to handle. May be use hping3 to send the ICMP packets and just use netmap to see how much packets it is able to receive in user space. It is pretty good. Once you get the netmap working, you can think of integrating your stack with netmap for a generic solution.

There are some good netmap videos available in ytube which explains the details of netmap.

from mtcp.

Thanks.

I also experienced netmap before and I could get 14.88 Mpps with it using
single core.
However, with larger number of NIC queues and larger payload sizes,
io_engine showed better performance.
For performance reason, we chose io_engine against netmap as the underlying
packet I/O solution.

DPDK is now emerging very fast and I think it comes with both performance
and generality.
We are thinking of DPDK as the next packet I/O library for mTCP, but we
will also discuss about the netmap.

Thanks,
EunYoung

2014-05-07 18:37 GMT+09:00 nginxsantos [email protected]:

Try this:

git clone https://code.google.com/p/netmap/

You can run without any drivers (the driver improves the performance). It
has some examples. You can just see how much PPS can the netmap able to
handle. May be use hping3 to send the ICMP packets and just use netmap to
see how much packets it is able to receive in user space. It is pretty
good. Once you get the netmap working, you can think of integrating your
stack with netmap for a generic solution.

There are some good netmap videos available in ytube which explains the
details of netmap.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1#issuecomment-42407279
.

from mtcp.