diff options
| author | Mao Zhongyi <maozy.fnst@cn.fujitsu.com> | 2017-12-25 10:54:12 +0800 |
|---|---|---|
| committer | Jason Wang <jasowang@redhat.com> | 2018-01-29 16:05:37 +0800 |
| commit | f449c9e549c2f93839a805ea9c7d66c9699794af (patch) | |
| tree | 51d6da09fe21a28241280d97d2bb3485ca311401 /net/colo.h | |
| parent | 9394133f8678ea02bbabab11c8e89e576ec8e9e8 (diff) | |
| download | qemu-f449c9e549c2f93839a805ea9c7d66c9699794af.tar.gz | |
colo: compare the packet based on the tcp sequence number
Packet size some time different or when network is busy.
Based on same payload size, but TCP protocol can not
guarantee send the same one packet in the same way,
like that:
We send this payload:
------------------------------
| header |1|2|3|4|5|6|7|8|9|0|
------------------------------
primary:
ppkt1:
----------------
| header |1|2|3|
----------------
ppkt2:
------------------------
| header |4|5|6|7|8|9|0|
------------------------
secondary:
spkt1:
------------------------------
| header |1|2|3|4|5|6|7|8|9|0|
------------------------------
In the original method, ppkt1 and ppkt2 are different in size and
spkt1, so they can't compare and trigger the checkpoint.
I have tested FTP get 200M and 1G file many times, I found that
the performance was less than 1% of the native.
Now I reconstructed the comparison of TCP packets based on the
TCP sequence number. first of all, ppkt1 and spkt1 have the same
starting sequence number, so they can compare, even though their
length is different. And then ppkt1 with a smaller payload length
is used as the comparison length, if the payload is same, send
out the ppkt1 and record the offset(the length of ppkt1 payload)
in spkt1. The next comparison, ppkt2 and spkt1 can be compared
from the recorded position of spkt1.
like that:
----------------
| header |1|2|3| ppkt1
---------|-----|
| |
---------v-----v--------------
| header |1|2|3|4|5|6|7|8|9|0| spkt1
---------------|\------------|
| \offset |
---------v-------------v
| header |4|5|6|7|8|9|0| ppkt2
------------------------
In this way, the performance can reach native 20% in my multiple
tests.
Cc: Zhang Chen <zhangckid@gmail.com>
Cc: Li Zhijian <lizhijian@cn.fujitsu.com>
Cc: Jason Wang <jasowang@redhat.com>
Signed-off-by: Mao Zhongyi <maozy.fnst@cn.fujitsu.com>
Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
Signed-off-by: Zhang Chen <zhangckid@gmail.com>
Reviewed-by: Zhang Chen <zhangckid@gmail.com>
Tested-by: Zhang Chen <zhangckid@gmail.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Diffstat (limited to 'net/colo.h')
| -rw-r--r-- | net/colo.h | 15 |
1 files changed, 15 insertions, 0 deletions
diff --git a/net/colo.h b/net/colo.h index 0658e869b4..da6c36dcf7 100644 --- a/net/colo.h +++ b/net/colo.h @@ -45,6 +45,15 @@ typedef struct Packet { int64_t creation_ms; /* Get vnet_hdr_len from filter */ uint32_t vnet_hdr_len; + uint32_t tcp_seq; /* sequence number */ + uint32_t tcp_ack; /* acknowledgement number */ + /* the sequence number of the last byte of the packet */ + uint32_t seq_end; + uint8_t header_size; /* the header length */ + uint16_t payload_size; /* the payload length */ + /* record the payload offset(the length that has been compared) */ + uint16_t offset; + uint8_t flags; /* Flags(aka Control bits) */ } Packet; typedef struct ConnectionKey { @@ -64,6 +73,12 @@ typedef struct Connection { /* flag to enqueue unprocessed_connections */ bool processing; uint8_t ip_proto; + /* record the sequence number that has been compared */ + uint32_t compare_seq; + /* the maximum of acknowledgement number in primary_list queue */ + uint32_t pack; + /* the maximum of acknowledgement number in secondary_list queue */ + uint32_t sack; /* offset = secondary_seq - primary_seq */ tcp_seq offset; /* |