Bad throughput for single tcp connections

Searching for "high latency file transfer" brings up a lot of interesting hits. Clearly, this is a problem that both the CompSci community and the commercial community has put thougth into.

A few commercial offerings that appear to fit the bill:

• FileCatalyst has products that can stream data over high-latency networks either using UDP or multiple TCP streams. They've got a lot of other features, too (on-the-fly compression, delta transfers, etc).

• The fasp file transfer "technology" from Aspera appears to fit the bill for what you're looking for, as well.

In the open-source world, the uftp project looks promising. You don't particularly need its multicast capabilities, but the basic idea of blasting out a file to receivers, receiving NAKs for missed blocks at the end of the transfer, and then blasting out the NAK'd blocks (lather, rinse, repeat) sounds like it would do what you need, since there's no ACK'ing (or NAK'ing) from the receiver until after the file transfer has completed once. Assuming the network is just latent, and not lossy, this might do what you need, too.

This problem was interesting as I've often wondered myself. I did a couple tests and found some interesting results. If I open up one connection to a server and wait 60 seconds it was invariably cleaned up(never got to 0.00/0/0). If I opened 100 connections, they too were cleaned up after 60 seconds. If I opened 101 connections I would start to see connections in the state you menitoned(that I've also seen before). And they appear to last roughly 120s or 2xMSL(which is 60) regardless of what fin_timeout is set to. I did some digging in the kernel source code and found what I believe is the 'reason'. There appears to be some code that tries to limit the amount of socket reaping that happens per 'cycle'. The cycle frequency itself is set on a scale based on HZ:

linux-source-2.6.38/include/net/inet_timewait_sock.h:
     35 #define INET_TWDR_RECYCLE_SLOTS_LOG     5
     36 #define INET_TWDR_RECYCLE_SLOTS         (1 << INET_TWDR_RECYCLE_SLOTS_LOG)
     37 
     38 /*
     39  * If time > 4sec, it is "slow" path, no recycling is required,
     40  * so that we select tick to get range about 4 seconds.
     41  */
     42 #if HZ <= 16 || HZ > 4096
     43 # error Unsupported: HZ <= 16 or HZ > 4096
     44 #elif HZ <= 32
     45 # define INET_TWDR_RECYCLE_TICK (5 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     46 #elif HZ <= 64
     47 # define INET_TWDR_RECYCLE_TICK (6 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     48 #elif HZ <= 128
     49 # define INET_TWDR_RECYCLE_TICK (7 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     50 #elif HZ <= 256
     51 # define INET_TWDR_RECYCLE_TICK (8 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     52 #elif HZ <= 512
     53 # define INET_TWDR_RECYCLE_TICK (9 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     54 #elif HZ <= 1024
     55 # define INET_TWDR_RECYCLE_TICK (10 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     56 #elif HZ <= 2048
     57 # define INET_TWDR_RECYCLE_TICK (11 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     58 #else
     59 # define INET_TWDR_RECYCLE_TICK (12 + 2 - INET_TWDR_RECYCLE_SLOTS_LOG)
     60 #endif
     61 
     62 /* TIME_WAIT reaping mechanism. */
     63 #define INET_TWDR_TWKILL_SLOTS  8 /* Please keep this a power of 2. */
The number of slots is also set here:
     65 #define INET_TWDR_TWKILL_QUOTA 100

In the actual timewait code you can see where it uses the quote to stop killing off TIME_WAIT connections if its already done too many:

linux-source-2.6.38/net/ipv4/inet_timewait_sock.c:
    213 static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
    214                                     const int slot)
    215 {
...
    240                 if (killed > INET_TWDR_TWKILL_QUOTA) {
    241                         ret = 1;
    242                         break;
    243                 }

Theres more information here on why HZ is set to what it is: http://kerneltrap.org/node/5411 But it isn't uncommon to increase it. I think however its usually more common to enable tw_reuse/recycling to get around this bucket/quota mechanism(which seems confusing to me now that I've read about it, increasing HZ would be a much safer and cleaner solution). I posted this as an answer but I think there could be more discussion here about what the 'right way' to fix it is. Thanks for the interesting question!