In my testing tuning the TCP stack had no significant benefit to SMB traffic. Other traffic like FTP does shows a benefit. It appears that SMB has its own control mechanism for sending traffic which is layered on top of TCP. Vista and up have SMB version 2 and 2.1(win7) they show some improvement because Microsoft has made the move from a block transfer protocol to something more like a stream.
I point this out as an example of an application that suffers from what I would consider a small amount of latency. -----Original Message----- From: Chris Cappuccio [mailto:[email protected]] Sent: Thursday, February 24, 2011 11:15 PM To: Jensen Tyler Cc: Jeff Cadwallader; Doug Hanks; Juniper-Nsp List Subject: Re: [j-nsp] Qfabric This sounds more like a configuration problem than a latency problem. Windows Vista/2008 and higher will auto-tune TCP window size to take advantage of available bandwidth, even if the latency creeps up. (So will all other modern operating systems at this point) You can always manually increase the default window size in older Windows servers and clients, through the registry. Adjusting the registry is a lot cheaper than deploying NYSE-euronext grade switches. But I guess if you first make a profit selling the switches, then make more installing them, it's a great deal for everyone, including me. I'll end up buying your old switches for 10% of the original sale price :) Jensen Tyler [[email protected]] wrote: > In my tests I have seen as much as a 30% drop in Windows file sharing > performance with 2 ms of latency vs <1ms. This was in a large radiology > application. Applications like FTP work without any issues. Some applications > are more sensitive(SMB). Low latency to me is measure in microseconds not > milliseconds(mostly layer 2). > > -----Original Message----- > From: [email protected] > [mailto:[email protected]] On Behalf Of Jeff Cadwallader > Sent: Thursday, February 24, 2011 11:31 AM > To: Doug Hanks > Cc: Juniper-Nsp List > Subject: Re: [j-nsp] Qfabric > > I deal with a lot of those issues also and usually when I ask what do they > mean by low latency the response comes back with sub-25ms. My data center is > all 1-2ms max on an aging platform. > > The other question I have is what happens to that entire logical device when > it fails in spectacular ways. > > I also agree that a sub-ms approach is needed in certain areas, however a > tiered approach has its advantages also. > > We are looking and evaluating replacing our aging platform in the data > center and will be following this closely. > > Jeff Cadwallader > > On Thu, Feb 24, 2011 at 11:15 AM, Doug Hanks <[email protected]> wrote: > > > A lot of our customers require low latency: financial, higher education, > > HPC environments and utility. > > > > Juniper has taken the time to solve more than just the low latency problem. > > We're trying to solve larger problems such as how do you manage an entire > > campus or data center as one logical device; that's able to scale; and > > delivers performance and low latency. > > > > Doug > > > > > _______________________________________________ > juniper-nsp mailing list [email protected] > https://puck.nether.net/mailman/listinfo/juniper-nsp > > _______________________________________________ > juniper-nsp mailing list [email protected] > https://puck.nether.net/mailman/listinfo/juniper-nsp -- the preceding comment is my own and in no way reflects the opinion of the Joint Chiefs of Staff _______________________________________________ juniper-nsp mailing list [email protected] https://puck.nether.net/mailman/listinfo/juniper-nsp
