Contact addresses for this AsymmetricThroughput case are:

• -- MarcoMarletta - marco.marletta@garr.it
• Mr. Monleon at ESTEC - Ricardo.Berto.Monleon@esa.int

Issue Description (MarcoMarletta)

ESA is performing some network performance test between their sites at:

• EsaEsrinSite, near Rome, Italy, connected to GARR with 32Mbps PVC over OC3 ATM link
• EsaEstecSite, in Netherlands, connected to SURFNet (1Gb/s to SURFnet's Leiden router)

They are experiencing asymmetric throughput, around 20Mbps from IT to NL, less than 3Mbps from NL to IT.

The application they use to test throughput is FTP (sigh) on Windows hosts (sigh!)

Hosts at EsaEstecSite: grid-medialab1.estec.esa.int [195.169.140.50] and grid-medialab4.estec.esa.int [195.169.140.53]
Hosts at EsaEsrinSite: esredi01.esa.int [193.204.228.68]

Current Status

The measurements taken on 22Apr2005 show that high TCP transfer rates can now be achieved at EsaEstecSite, in both directions. One bottleneck that was found is that on Windows, it doesn't seem to be possible to set a system-wide default for the transmit-buffer size for all TCP sockets. Thus, even when large windows are possible, applications cannot use them unless they use specific APIs such as setsockopt(...,SO_SNDBUF,...). Using the iperf tool with the right options, ChrisWelti was able to overcome this issue.

One open question concerns the sometimes very low rates that were observed for TCP transfers from ESTEC to other sites. These rates were so low that they could hardly be explained by TCP window/buffer issues. We could not reproduce them lately. It is quite possible that, during the long time this issue was under investigation, other things in the network have been improved.

Another open question is: why the asymmetry, if all Windows systems have the same problem? In order to answer this, we'd have to know how the tests were done. For example, if all initial testing was done with an FTP client at EsaEstecSite and an FTP server at EsaEsrinSite, a satisfactory explanation would be that the FTP server did something special (e.g. setsockopt) to obtain larger TCP transmission windows, while the client uses the small default.

Summary Results

Path	Test type	Data rate	Sender's Buffer
ESRIN -> ESTEC	FTP	20Mbps	64k?
ESTEC -> ESRIN	FTP	3Mbps	64k?
ESTEC -> SWITCH	NDT	20Mbps	64k
SWITCH -> ESTEC	NDT	65Mbps	8M
ESTEC -> SWITCH	iperf	90Mbps	8M
SWITCH -> ESTEC	iperf	70Mbps	8M

Log Entries (inverse chronological order)

Thanks to remote access to a laptop in the ESTEC LAN we have some new results. However, the results seem to suggest that there is no performance problem at the current location where the laptop is attached. The laptop has a 100Mbit/s full-duplex access-link and we can achieve around 90Mbit/s upload speed to ndt.switch.ch and around 65-70Mbit/s download speed from ndt.switch.ch (I suppose this is due to other traffic in that specific vlan the laptop is in).

What we observed is that you can not set a default send window for TCP so you will end up with a 64K default window which will lead to around 20Mbit/s upload speed limitation as observed in other tests before. However when using an application which uses setsockopt to set the buffers manually to 512K (like in this case iperf) we were able to push it to 90Mbit/s.

It is still unclear to me why the test which was conducted by Marco on Monday only resulted in around 200Kbit/s upload speed and 18.81Mbit/s; Talking with Marco about this it seems that nothing has changed since then.

-- ChrisWelti - 22 Apr 2005

---

For new measurements (Web100 NDT) at EsaEstecSite, please check out 19Apr2005.

-- SimonLeinen - 19 Apr 2005

---

In a first attempt to isolate the problem I repeated download and upload transfers from another host connected to GARR backbone (incidentally on the same LAN of pert issue1) instead of behind the ESRIN access. Result was:

NL-IT, 40MB and 4MB test files,  2320Kbps
IT-NL, 40MB and 4MB test files, 16000Kbps

I am currently collecting troubleshooting data, as remote access to machines at both EsaEsrinSite and EsaEstecSite (now I have only FTP access to machines at ESTEC).

-- MarcoMarletta - 29 Nov 2004

---

At first sight, something is occurring at EsaEstecSite, but who knows..

-- PavelCimbal - 30 Nov 2004

Issue Tracker Reference

This is Issue #3 on the Pilot PERT Roundup Tracker (PRI):

http://issues.pert.geant2.net:8080/pert/issue3

-- SimonLeinen - 01 Dec 2004