UNSOLVED

shaka1

updated

20 years ago

S

shaka1

3 Posts

0

1566

April 16th, 2007 05:00

NS350 CIFs performance

I have NS350. What is maximum performance I can get from it ? MB/SEC !!
  • shaka1

    3 Posts

    385

    0

    Posted April 17th, 2007 01:00

    Please share with me what is your CIFS Share maximum performance ???
  • dynamox

    11 Legend

    •

    20419 Posts

    •

    87439 Points

    385

    0

    Posted April 18th, 2007 17:00

    i will depend on your network/celerra configuration. I am not doing any port trunking on the network switch, so my celerra interfaces are connected to a 1G network switch, while my clients are all over the company campus and are connected to 100MB switches. On average we are seeing 10MB/s throughput to users desktops.
  • Duane_Olson

    16 Posts

    385

    0

    Posted May 11th, 2007 09:00

    THe maximum theoretical limit of GigE is 125MB/sec. (1000/8=125). In CIFS, the CIFS protocol consumes approximately 15% of the available bandwidth to do non-i/o type work...authentication, etc. This leaves you with 106MB/sec. THere is an additional 15% that is taken away by the network itself for things sucha as ARP broadcasts, etc. Leaving about 90MB/sec. In the real world, it is hard to see these numbers, but on a single NIC in a host, these are the maximums you can expect. The datamover is capable of 4x this number if all 4 interfaces are used. LACP and etherchannel are both supported, but this only add more lanes on the highway, it does not increase the speed limit on the network. A single NIC can still only push about 90MB/sec of traffic. Realsize also that this is full duplex(unless you are in hubs which only support 1/2 duplex). The 90MB is read, and there is an additional 90MB of write that can occur simultaneously. You cannot "borrow" from the other however. THe max is 90MB of write, and 90MB of read per NIC.
  • mssqadri

    8 Posts

    385

    0

    Posted June 3rd, 2010 23:00

    Hi,

    We were also having performance bottleneck while reading data from Celerra. After some investigation and consultation with EMC, we have changed fastRTO setting on data-movers from default value of "0" to "2" and now getting almost full throughput. Before this change we were getting very low throughput (<10Mbps) while reading from any Celerra machine.

    In our scenario, Celerra is connected to 3Com Core Switch 8810 via aggregated links of 2x1Gbps and clients are connected through edge switch on 100Mbps ports. Now clients are getting 90 to 100 Mbps.

    Can you share your fastRTO current settings?

    Regards,

    Saleem

  • IanSchorr

    117 Posts

    385

    0

    Posted June 6th, 2010 22:00

    Hi Saleem,

    You posted this question in response to about 10 old threads (this one, for example, is 3 years old).  I'm going to pick this thread at random to reply to.  I'm not sure if you're trying to help, or if you're looking for more information about fastRTO.

    A couple things:

    1) Do you know why tcp.fastRTO value of "2" was chosen?  fastRTO is a boolean param, either on or off.  A0 is off.  values of 1, 2, 3, 4, etc all mean "on".

    2) If you've been asked to enable this param, it means that you have significant packet drops on your network, and resolving those issues are going to give you best performance.

    Depending on your systems' NFS or CIFS workload and the packet drop pattern that's happening, TCP may or may not be able to "detect" when packet drops occur.  NFS and CIFS (and iSCSI) are block-oriented (as opposed to stream-oriented) protocols, which means they tend to request data in rather small chunks.  CIFS on Windows 2003 and below tends to request only one block at a time, and all this means that it's significantly more likely that TCP won't be able to "quickly" detect that packets have been dropped and recover from them.

    When TCP can't detect this, packets are only retransmitted (and data transfer continue) after a TCP timeout - the RTO (Retransmission TimeOut).

    By default, Celerra retransmits after 1.2-1.6s, which is per TCP standards.  Setting the param causes Celerra to retransmit after .4-.6s, which technically breaks the rules, but is virtually always acceptable (particularly within datacenters) and many modern OSes do this by default.  Setting the param means we recover faster from packet drops, but it means that you're still having .4-.6s periods, probably frequently, where no data is being transferred.

    That's good that it helped resolve much of your performance loss.  It's also a param that is generally very safe to set and will help us recover faster in networks with packet drops.  But it's not a fix-all, and actually discourages people from addressing the real problem and getting maximum performance (packet drops), and rarely is a problem that affects/would affect only Celerra.

    These kinds of packet drops are due to network congestion 99% of the time.  And network congestion is generally due to:

    1) Underpowered switch equipment/bandwidth (the switch itself is unable to handle the amount of traffic being passed through it)

    2) significantly oversubscribed/overwhelmed network link (for example, trying to stuff 1Gbps of data down a 100Mbps pipe, and/or a large number of systems attempting to talk at high speeds over a single trunk of similar speed)

    3) insufficient network buffers (including unnecessarily small port buffers.  Often the result of poor QoS configuration)

    It sounds like you may be suffering from #1 and #2.

    These kinds of problems will eventually be addressed by "Lossless Ethernet", which enables end-to-end Ethernet flow control that virtually eliminates these kinds of drops.  That's a few years out - very few vendors (other than Cisco) support this yet, and there are a few kinks to work out.  But avoiding packet drops is essential for good performance on an Ethernet-based storage network.