Using netstat and dropwatch to observe packet loss on Linux servers
Anyone that is running a modern Operating System is most likely utilizing TCP/IP to send and receive data. Modern TCP/IP stacks are somewhat complex and have a slew of tunables to control their behavior. The choice of when and when not to tune is not always super clear cut, since documentation and the advice of various network “experts” doesn’t always jive.
When I’m looking into performance problems that are network related one of the first things I review is the netstat “-s” output:
$ netstat -s
Ip: 25030820 total packets received 269 with invalid addresses 0 forwarded 0 incoming packets discarded 21629075 incoming packets delivered 21110503 requests sent out Icmp: 12814 ICMP messages received 0 input ICMP message failed. ICMP input histogram: destination unreachable: 2 echo requests: 12809 echo replies: 3 12834 ICMP messages sent 0 ICMP messages failed ICMP output histogram: destination unreachable: 22 echo request: 3 echo replies: 12809 IcmpMsg: InType0: 3 InType3: 2 InType8: 12809 OutType0: 12809 OutType3: 22 OutType8: 3 Tcp: 138062 active connections openings 1440632 passive connection openings 7 failed connection attempts 2262 connection resets received 8 connections established 12225207 segments received 10785279 segments send out 10269 segments retransmited 0 bad segments received. 69115 resets sent Udp: 553643 packets received 22 packets to unknown port received. 0 packet receive errors 6911684 packets sent UdpLite: TcpExt: 33773 invalid SYN cookies received 154132 TCP sockets finished time wait in fast timer 6 time wait sockets recycled by time stamp 72284 delayed acks sent 3 delayed acks further delayed because of locked socket Quick ack mode was activated 269 times 3359 packets directly queued to recvmsg prequeue. 2592713 packets directly received from backlog 4021 packets directly received from prequeue 3557638 packets header predicted 1732 packets header predicted and directly queued to user 1939991 acknowledgments not containing data received 3179859 predicted acknowledgments 1631 times recovered from packet loss due to SACK data Detected reordering 1034 times using FACK Detected reordering 1007 times using SACK Detected reordering 622 times using time stamp 1557 congestion windows fully recovered 4236 congestion windows partially recovered using Hoe heuristic 299 congestion windows recovered after partial ack 2 TCP data loss events 5 timeouts after SACK recovery 5 timeouts in loss state 2511 fast retransmits 2025 forward retransmits 88 retransmits in slow start 5518 other TCP timeouts 295 DSACKs sent for old packets 35 DSACKs sent for out of order packets 251 DSACKs received 25247 connections reset due to unexpected data 2248 connections reset due to early user close 6 connections aborted due to timeout TCPSACKDiscard: 2707 TCPDSACKIgnoredOld: 65 TCPDSACKIgnoredNoUndo: 12 TCPSackShifted: 4176 TCPSackMerged: 2301 TCPSackShiftFallback: 98834 IpExt: InMcastPkts: 2 OutMcastPkts: 3390453 InBcastPkts: 8837402 InOctets: 5156017179 OutOctets: 2509510134 InMcastOctets: 80 OutMcastOctets: 135618120 InBcastOctets: 2127986990
The netstat output contains a slew of data you can be used to see how much data your host is processing, if it’s accepting and processing data efficiently and if the buffers that link the various layers (Ethernet -> IP -> TCP -> APP) are working optimally.
When I build new Linux machines via kickstart, I make sure my profile contains the ktune package. That is all the tuning I do to start, unless an application or database requires a specific setting (think large pages and SysV IPC settings for Oracle).
Once I’ve met with an application resource and a business analyst, I like to pound the application with a representative benchmark and compare the system performance before and after the stress test was run. By comparing the before and after results I can see where exactly the system is choking (this is very rare), or if the application needs to be modified to accommodate additional load. If the application is a standard TCP/IP based application that utilizes HTTP, I’ll typically turn to siege and iPerf to stress my applications and systems.
If during load-testing I notice that data is being dropped in one or more queues, I’ll fire up dropwatch to observe where in the TCP/IP stack data is being dropped:
$ dropwatch -l kas
Initalizing kallsyms db dropwatch> start Enabling monitoring... Kernel monitoring activated. Issue Ctrl-C to stop monitoring 1 drops at netlink_sendskb+14d (0xffffffff813df30e) 1 drops at ip_rcv_finish+32e (0xffffffff813f0c93) 4 drops at ip_local_deliver+291 (0xffffffff813f12d7) 64 drops at unix_stream_recvmsg+44a (0xffffffff81440fb9) 32 drops at ip_local_deliver+291 (0xffffffff813f12d7) 23 drops at unix_stream_recvmsg+44a (0xffffffff81440fb9) 1 drops at ip_rcv_finish+32e (0xffffffff813f0c93) 4 drops at .brk.dmi_alloc+1e60bd47 (0xffffffffa045fd47) 2 drops at skb_queue_purge+60 (0xffffffff813b6542) 64 drops at unix_stream_recvmsg+44a (0xffffffff81440fb9)
This allows you to see if data is being dropped at the link layer, the IP layer, the UDP/TCP layer or the application layer. If the drops are occurring somewhere in TCP/IP (i.e. inside the kernel) I will review the kernel documentation and source code to see what occurs at the specific areas of the kernel listed in the dropwatch output, and find the sysctl values that control the sizes of the buffers at that layer (some are dynamic, some are fixed).
Tuning applications to perform optimally has filled dozens and dozens of books, and it’s a fine art that you learn from seeing problems erupt in the field. It also helps to know how to intepret all the values in the netstat output, and I cannot recommend TCP/IP volume I, TCP/IP volume II and TCP/IP volume III enough! Everyone who runs an IP connected system should be required to read these before they are allowed access to the system. :)