Pages

Tuesday, June 09, 2009

How to speed up data transfer between servers

This article contains values as an example of a specific case. Your servers may have different configuration. Please use accordingly.

Over a network, although the network bandwidth could be 10 GigaBits, yet, if the servers have
1 GigE Ethernet network cards (1 GigaBit), the overall file-transfer
throughput is 1 GigaBits / sec, which means (1024*1024*1024 / 8) bytes
per sec = 125 MB/sec. This is the maximum throughput available on the network.


  • For convenience, the 2 esrvers we need to improve data transfer speeds between - we'd call one as source, & other target. Now, if we ping from a source to target server, we get about 42 ms of turnaround time.






  • This gives us maximum amount of data at any point of time possible on this network as (125 MB / s) * (0.042) = 5.250 MB . However, the default network level settings on the servers don’t let this much bandwith to be used. We may verify it with “ndd & route” commands (only SA can run these commands):



  • If you have root access to any locale Linux box, you may try "ndd get ..." & "route print" commands, otherwise ask the SA to provide you the outputs, to find out the current settings.



  • The **recv** & **xmit** (receive & transmit) values are the default ones, used by the server & maximum values are defined by **cwnd** & **max_buf. e.g. The 49152 value indicates that this server (or any RMAN job) would send / receive, by default, 48 KB of traffic only. Moreover, if someone is manually writing a program, may use the maximum of 1048576 bytes
    = 1 MB, by using the **max** variables)



  • Now, our purpose is to be able to use the whole 5.25 MB that can be used on the current
    connection. However, the allowed values for these 4 variables, should be a power of 2. Hence, we would pick the next higher power of 2, which is 8 MB (=2 ^ 3). We would need to set the **max** limit variables also equal to 8 MB.



  • DBA should ask SA to run these commands for setting the values, with “set” option, instead of “get”. E.g. “sudo ndd -set /dev/tcp tcp_recv_hiwat 8388608”



  • The “ndd” commands get / set values for ALL connections. We may not want to do that everytime,
    because if a non Netbackup / media server (say, a windows client) connects to this server, which has a lower capacity of being able to handle traffic (say, 1 MB), then the 5 MB packets would overwhelm the connection. Hence,
    to avoid this scenario, we should use “route” commands to set connections settings between specific IP’s , which don’t work the same as “ndd”.



  • It’s always safer to use “route” settings instead of
    “ndd”
    excepttcp_max_buf”. Here are some screenshots (please esarch internet to see more detailed syntax of route / ndd commands) :






  • The “sendpipe” option is same as “tcp_xmit_hiwat” of ndd & recvpipe is “tcp_recv_hiwat”.
    The “ssthresh” is same as “tcp_cwnd_max”. However, the “tcp_max_buf” still needs to be set through “ndd”.



  • With "route" settings, all these settings will be wiped out & will go back to default whenever a server is rebooted, so changes are temporary till next boot. With "ndd', it may be made permanent, which may be unnecessary, unless it's between some DB & RMAN backup server (viz. Netbackup) where there's a regular transfer of data.



  • "ndd" uses get / set options to show or edit values, whereas "route" uses print / add / delete options.



  • Now, having understood all the above concepts, when a DBA / SA is ready to start working on their specific environment of servers, you may follow this :

    1) 1st we need to get the current values from the SA, of the 4 parameters (to be executed from source / target server),
    a. ndd –get /dev/tcp tcp_recv_hiwat
    b. ndd –get /dev/tcp tcp_xmit_hiwat
    c. ndd –get /dev/tcp tcp_max_buf
    d. ndd –get /dev/tcp tcp_cwnd_max
    e. route get

    2) For “route get” output, these parameters are named sendpipe, recvpipe, ssthresh (there’s no equivalent of tcp_
    max_buf here – it can be set only through ndd command). “ndd” sets the parameters universally, for all IP’s, whereas “route” can be used to set between specific IP’s & temporarily.

    3) Along with this we need to know the server Ethernet card types (on both source & target servers) , viz. 1 GigE, 10GigE ; AND the network bandwidth , viz. a 10 Gbps line, 24 Mbps line etc.

    4) Then we need to make calculations based on the above network bandwidth values, taking into the consideration, an example calculation on my Blog - http://databaseoracle.blogspot.com/2009/06/how-to-speed-up-data-transfer-between.html .

    5) This way, we’d arrive at a number, which is the total bandwidth that CAN BE available between the 2 servers under consideration. We’ll call this number (if it is a power of 2) or the next power of 2, MAX-BANDWIDTH

    6) Then we request the SA to set the above parameters to new values (use route command, instead of ndd , to set these values temporarily) :
    a. route add –host … -sendpipe MAX-BANDWIDTH -recvpipe MAX-BANDWIDTH -ssthresh MAX-BANDWIDTH
    b. ndd –set /dev/tcp tcp_cwnd_max MAX-BANDWIDTH (this parameter is available only through ndd & SA may decide whether to set it or not, based on implications) .

    7) If the SA is responding in minutes, this whole calculation would take about 10-20 minutes.

    8) Please let me know in case of any doubts & your experience when you try these modifications.

    9) Happy transferring lightning-fast data !