Comentários (1)

1 rohanlon comentou às Link permanente

Seb, <br /> What is the comparative latency hit of FEC vs the Cut Thru switching? <br /> Why not just use the CRC check as it was intended? <div>&nbsp;</div> Maybe you could write an article on the latency characteristics of a SAN. This could include the impact of variable latency and jitter vs consistent latency. <div>&nbsp;</div> So, my question is, WHO CARES? <br /> Cisco gear is 15-20uS (any port to any port) and Brocade which has different latency depending on whether its on the same asic or different asic / different linecard (big variation). Surely as a troubleshooter, you want to know what your latency should be through the box? <div>&nbsp;</div> SAN latency is a funny thing. If you are accessing a disk with a cache miss, you are likely into hundreds of milliseconds which the disk is accessed. If you have SSD DAS, it's microseconds (but no SAN here) <div>&nbsp;</div> If you know that corrupt frames are dropped at ingress, then you don't have to waste time finding the source (the switch will tell you) and the corrupt frames are not clogging up the remote HBA or disk controller whilst it instigates error recovery for the corrupt frame. If this is a heavily used array port, you really don't want this big shared resource wasting time &amp; resources dropping the frame. <br /> If this is an HBA, it could be the instigator of slow drain issues across the fabric. <div>&nbsp;</div> I work for Cisco, and I was around in the days when Kalpana invented the first cut-thru switch (Cisco bought them in 1994). There was a lot of hype back then, but nowadays it's not very interesting because the ASIC technology has moved on such that the CRC check can be done at wire rate. Ultra low latency switches exist now for niche applications and they are typically combined with specialized protocol stacks that remove the latencies that are inherent to IP. <div>&nbsp;</div> So, why not implement the CRC check as it was intended? Drop the frame at ingress and avoid all this hassle when you have to troubleshoot the network. <div>&nbsp;</div> Like I say, I work for Cisco, but I'm genuinely intrigued why you would not drop bad frames before they enter the fabric. <div>&nbsp;</div> <div>&nbsp;</div>