Start a Conversation

Unsolved

This post is more than 5 years old

4810

April 28th, 2013 21:00

Celerra TCP window size

Hello!

I have Celerra CX4-960. And have a problem with speed. It's only 48mbits/sec. I analyzed traffic with Wireshark. TCP Window only 17KB. How i can increase tcp window size? Which parameter i would change to it? i found this parameters: RCP TCP window; tcp sndcwnd. But no one of their didnt help to change tcp window size.

4 Operator

 • 

8.6K Posts

April 29th, 2013 02:00

Hi,

what protocols do you specifically mean ?

a CX4-960 isnt a Celerra - it’s a Clariion.

7 Posts

April 29th, 2013 02:00

Sorry! Yes, it's clariion.

I mean CIFS and NFS. Some customers connects with NFS, some with CIFS

4 Operator

 • 

8.6K Posts

April 29th, 2013 03:00

Well if it is just a Clariion then it doesn’t speak NFS or CIFS – just FC or ISCSI

You would either have a Celerra gateway in front of it or some Windows / Unix servers

4 Operator

 • 

8.6K Posts

April 29th, 2013 04:00

BTW these parameters wont help you

the rcp param is only effective for replication with Replicator or nas_copy anyway

On the Celerra we automatically negotiate to the largest possible TCP Window size that the client supports.

If you have performance problems - its not because of tcp window size - that typically only make a difference on WAN connections.

I suggest to open a service request with EMC customer service.

They can also analyze network trace to see if its really a network problem.


4 Operator

 • 

8.6K Posts

April 29th, 2013 04:00

Well – if it is a Celerra than I suggest to first try setting the fastRTO parameter before playing with tcp window size manually

Do a search here on the forum for fastRTO – there should be a number of posts from customers where it helped

7 Posts

April 29th, 2013 04:00

I'm a new ingeneer in a company. I dont know this system will, like other ingeneers. Maybe we have celerra gateway. But i know, that we create nfs and cifs servers and shares.

7 Posts

April 29th, 2013 05:00

Oh, thank you!

I found posts about fastRTO. THere people said to change fastRTO value from 0 to 2. In a manual to system i read that only 2 values: 0 or 1. I think that 0 is OFF, all other values (1,2,3 or other) is mean ON? Right?

4 Operator

 • 

8.6K Posts

April 29th, 2013 05:00

Yes – setting it to 1 means fastRTO on

7 Posts

April 30th, 2013 01:00

We are did connect 2 PC on windows XP with same network. In registry we changed tcp windows size (about 202KB), tcp1323options (window scaling and timestamps) and use tcp optimizer utility to change all parameters to optimal. After that we get 10-12 MB/sec data transfer.

Does your system have same options (tcp window size, tcp1323 options & others)?

7 Posts

April 30th, 2013 01:00

We changed fastRTO. Unfortunately, unhelpfully.

We examinated connection speed. When we send data to storage speed about 5-6MB/sec, tcp window 251byte(!), when we copy data from storage speed is 17kB and speed about 2MB/sec.

How i could open service request?

4 Operator

 • 

8.6K Posts

April 30th, 2013 06:00

Use support.emc.com or live chat or telephone to open a SR (assuming you have a support contract)

4 Operator

 • 

8.6K Posts

April 30th, 2013 06:00

Sounds like something on the network is wrong

46 Posts

May 4th, 2013 19:00

What we've had to do is *disable* large window sizes on the clients.  We found that the Celerra was overrunning the host switch ports, the switch buffer would fill and would drop packets before the tcp window sizes could scale back effectively (clients would advertise a fairly large window size, which instead of starting slowly the Celerra used that to start).  Unfortunately as per TCP spec when you drop (from memory I think 3x) packets in a row the sender will backoff for 2 seconds, which is an eternity.  Turning up the wireshark throughput resolution you'd see a surging pattern where you would get lots of data for a very short time and then nothing.  With tcp scaling disabled on a Linux box over NFS we went from 10-15MB/sec to >100MB/sec.  The celerra also has a parameter sndcwnd, which is the max window size the celerra will use for all connections.  We didn't want to affect everybody just the high bandwidth ones that were having problems, so we turned of scaling on those particular hosts which seemed to fix our throughput problems. 

Our issue came from the differing network rates, our Celerra has 10gig and most of our clients are 1gig, the 10gig would just overrun the switch buffers (even on our Cisco nexus line with it's larger buffer capacity).  FastRTO helped with reducing the impact when packets were dropped but disabling scaling gave us wire rate speeds.

4 Operator

 • 

8.6K Posts

May 5th, 2013 08:00

Yep - larger isn't always better - especially when the network can't keep up

Packet loss has a serious impact on performance

Good test is to run ttcp between client and data mover - it doesn't use disk and higher protocols

7 Posts

June 16th, 2013 21:00

Hm... How we can turn on window scaling? Our customers is far away from our system. and between their we have a lot of network devices. Maybe we can increase window size to 60-100kb? We run test connection without clariion. result: with windows size 64kb we reach 90MB/s

No Events found!

Top