If jumbo frames are enabled only on uplinks but not all the way between two systems, then the end systems won't take any advantage of jumbo frames. There is no drawbacks of jumbo frames as such as far as I know, but some pitfalls.
Jumbo frames are any frames bigger than standard Ethernet frames (1518 bytes of user-visible part). And some platforms implement jambo frames as big as 9216 bytes (Cat 6500), while others (e.g. Cat 2950) are limited to baby-jumbo of 1530 bytes. So when you enable jumbo frames you must be sure that the size of jumbo frame is consistent across all your systems including servers and client PC's connected to your network.
Another pitfall is that if you enable jumbo frame on any IP-layer interface this will automatically change IP MTU. If you're running OSPF and jumbo frames are enabled only on some systems connected to a subnet but not on other systems from the same subnet OSPF adjacencies will not form until you specify 'ip mtu 1500' on jumbo-enabled systems. As soon as you do this, effect of jumbo frames for IP traffic will be void (but it might still be necessary for things like MPLS). So be sure that systems on common subnet have same MTU.
Routing problem is easy to detect, more general problem is that travelling across L2-only path there is no way for switches to send 'ICMP Fragmentation required' if packet is large then next interface MTU. This will break PMTU Discovery and since most applications usually sent packets at max MTU and DF-bit set, there will be timeouts. So again, consistent MTU across whole L2 path is important.
By the way, if your servers and PC's are not connected via jumbo-enabled links then you unlikely see any difference by enabling jumbo frames on the uplinks because both 6500 and 3550 catalysts are capable of wirespeed performance. The only time when it makes sense to enable jumbo frames only on the core links is when you need some non-IP headers to encapsulate your max-sized IP packets (MPLS is one such example).
As for benefits - servers running (very) heavy traffic applications (think full-feed USENET server with multiple fast peerings) may benefit from sending large portion of data in each packet, so for the same amount of data they need less number of packets. Destination system will have to handle less interrupts and overall performance may increase.
Hope this helps.
