<div dir="ltr"><div>Hello,<br><br></div>vmstat seems pretty boring. Certainly nothing going to swap.<br><br><span style="font-family:monospace,monospace">root@sanbox:/root# vmstat<br> kthr      memory            page            disk          faults      cpu<br> r b w   swap  free  re  mf pi po fr de sr po ro s0 s2   in   sy   cs us sy id<br> 0 0 0 34631632 30728068 175 215 0 0 0 0 963 275 4 6 140 3301 796 6681 0  1 99</span><br><br><br><div>Here is the "taskq_dispatch_ent" output from "lockstat -s 5 -kWP sleep 30" during the "fast" write operation.<br></div><div><span style="font-family:monospace,monospace">-------------------------------------------------------------------------------<br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>50934   3%  79% 0.00     3437 0xffffff093145ba40     taskq_dispatch_ent<br><br>      nsec ------ Time Distribution ------ count     Stack<br>       128 |                               7         spa_taskq_dispatch_ent<br>       256 |@@                             4333      zio_taskq_dispatch<br>       512 |@@                             3863      zio_issue_async<br>      1024 |@@@@@                          9717      zio_execute<br>      2048 |@@@@@@@@@                      15904<br>      4096 |@@@@                           7595<br>      8192 |@@                             4498<br>     16384 |@                              2662<br>     32768 |@                              1886<br>     65536 |                               434<br>    131072 |                               34<br>    262144 |                               1<br>-------------------------------------------------------------------------------</span><br><br><br><br></div><div>However, the truly "broken" function is a read operation:<br><br></div><div>Top lock 1st try:<br><span style="font-family:monospace,monospace">-------------------------------------------------------------------------------<br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>  474  15%  15% 0.00     7031 0xffffff093145b6f8     cv_wait<br><br>      nsec ------ Time Distribution ------ count     Stack<br>       256 |@                              29        taskq_thread_wait<br>       512 |@@@@@@                         100       taskq_thread<br>      1024 |@@@@                           72        thread_start<br>      2048 |@@@@                           69<br>      4096 |@@@                            51<br>      8192 |@@                             47<br>     16384 |@@                             44<br>     32768 |@@                             32<br>     65536 |@                              25<br>    131072 |                               5<br>-------------------------------------------------------------------------------</span><br><br><br></div><div>Top lock 2nd try:<br></div><div><span style="font-family:monospace,monospace">-------------------------------------------------------------------------------<br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>  174  39%  39% 0.00   103909 0xffffff0943f116a0     dmu_zfetch_find<br><br>      nsec ------ Time Distribution ------ count     Stack<br>      2048 |                               2         dmu_zfetch<br>      4096 |                               3         dbuf_read<br>      8192 |                               4         dmu_buf_hold_array_by_dnode<br>     16384 |                               3         dmu_buf_hold_array<br>     32768 |@                              7<br>     65536 |@@                             14<br>    131072 |@@@@@@@@@@@@@@@@@@@@           116<br>    262144 |@@@                            19<br>    524288 |                               4<br>   1048576 |                               2<br>-------------------------------------------------------------------------------</span><br><br></div><div>Top lock 3rd try:<br><br><span style="font-family:monospace,monospace">-------------------------------------------------------------------------------<br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>  283  55%  55% 0.00    94602 0xffffff0943ff5a68     dmu_zfetch_find<br><br>      nsec ------ Time Distribution ------ count     Stack<br>       512 |                               1         dmu_zfetch<br>      1024 |                               1         dbuf_read<br>      2048 |                               0         dmu_buf_hold_array_by_dnode<br>      4096 |                               5         dmu_buf_hold_array<br>      8192 |                               2<br>     16384 |                               7<br>     32768 |                               4<br>     65536 |@@@                            33<br>    131072 |@@@@@@@@@@@@@@@@@@@@           198<br>    262144 |@@                             27<br>    524288 |                               2<br>   1048576 |                               3<br>-------------------------------------------------------------------------------</span><br><br></div><div><br></div><div>As for the MTU question- setting the MTU to 9000 makes read operations grind almost to a halt at 5MB/s transfer rate.<br><br></div><div>-Warren V<br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Mar 2, 2015 at 11:30 AM, Garrett D'Amore <span dir="ltr"><<a href="mailto:garrett@damore.org" target="_blank">garrett@damore.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word">Here’s a theory.  You are using small (relatively) MTUs (3000 is less than the smallest ZFS block size.)  So, when you go multipathing this way, might a single upper layer transaction (ZFS block transfer request, or for that matter COMSTAR block request) get routed over different paths.  This sounds like a potentially pathological condition to me.<div><br></div><div>What happens if you increase the MTU to 9000?  Have you tried it?  I’m sort of thinking that this will permit each transaction to be issued in a single IP frame, which may alleviate certain tragic code paths.  (That said, I’m not sure how aware COMSTAR is of the IP MTU.  If it is ignorant, then it shouldn’t matter *that* much, since TCP should do the right thing here and a single TCP stream should stick to a single underlying NIC.  But if COMSTAR is aware of the MTU, it may do some really screwball things as it tries to break requests up into single frames.)</div><div><br></div><div>Your read spin really looks like only about 22 msec of wait out of a total run of 30 sec.  (That’s not *great*, but neither does it sound tragic.)  Your write  is interesting because that looks like it is going a wildly different path.  You should be aware that the locks you see are *not* necessarily related in call order, but rather are ordered by instance count.  The write code path hitting the task_thread as hard as it does is really, really weird.  Something is pounding on a taskq lock super hard.  The number of taskq_dispatch_ent calls is interesting here.  I’m starting to wonder if it’s something as stupid as a spin where if the taskq is “full” (max size reached), a caller just is spinning trying to dispatch jobs to the taskq.  </div><div><br></div><div>The taskq_dispatch_ent code is super simple, and it should be almost impossible to have contention on that lock — barring a thread spinning hard on taskq_dispatch (or taskq_dispatch_ent as I think is happening here).  Looking at the various call sites, there are places in both COMSTAR (iscsit) and in ZFS where this could be coming from.  To know which, we really need to have the back trace associated. </div><div><br></div><div>lockstat can give this — try giving “-s 5” to give a short backtrace from this, that will probably give us a little more info about the guilty caller. :-)</div><div><br></div><div><span style="white-space:pre-wrap">       </span>- Garrett</div><div><br><div><blockquote type="cite"><div><div class="h5"><div>On Mar 2, 2015, at 11:07 AM, W Verb via illumos-developer <<a href="mailto:developer@lists.illumos.org" target="_blank">developer@lists.illumos.org</a>> wrote:</div><br></div></div><div><div dir="ltr"><div><div class="h5"><div><div><div>Hello all,<br></div>I am not using layer 2 flow control. The switch carries line-rate 10G traffic without error.<br><br></div>I think I have found the issue via lockstat. The first lockstat is taken during a multipath read:<br><br><br></div>lockstat -kWP sleep 30<br><div><div><div><br>Adaptive mutex spin: 21331 events in 30.020 seconds (711 events/sec)<br><br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>-------------------------------------------------------------------------------<br> 9306  44%  44% 0.00     1557 htable_mutex+0x370     htable_release<br> 6307  23%  68% 0.00     1207 htable_mutex+0x108     htable_lookup<br>  596   7%  75% 0.00     4100 0xffffff0931705188     cv_wait<br>  349   5%  80% 0.00     4437 0xffffff0931705188     taskq_thread<br>  704   2%  82% 0.00      995 0xffffff0935de3c50     dbuf_create<br><br></div><div>The hash table being read here I would guess is the tcp connection hash table.<br></div><div><br></div><div>When lockstat is run during a multipath write operation, I get:<br><br>Adaptive mutex spin: 1097341 events in 30.016 seconds (36558 events/sec)<br><br>Count indv cuml rcnt     nsec Hottest Lock           Caller<br>-------------------------------------------------------------------------------<br>210752  28%  28% 0.00     4781 0xffffff0931705188     taskq_thread<br>174471  22%  50% 0.00     4476 0xffffff0931705188     cv_wait<br>127183  10%  61% 0.00     2871 0xffffff096f29b510     zio_notify_parent<br>176066  10%  70% 0.00     1922 0xffffff0931705188     taskq_dispatch_ent<br>105134   9%  80% 0.00     3110 0xffffff096ffdbf10     zio_remove_child<br>67512   4%  83% 0.00     1938 0xffffff096f3db4b0     zio_add_child<br>45736   3%  86% 0.00     2239 0xffffff0935de3c50     dbuf_destroy<br>27781   3%  89% 0.00     3416 0xffffff0935de3c50     dbuf_create<br>38536   2%  91% 0.00     2122 0xffffff0935de3b70     dnode_rele<br>27841   2%  93% 0.00     2423 0xffffff0935de3b70     dnode_diduse_space<br>19020   2%  95% 0.00     3046 0xffffff09d9e305e0     dbuf_rele<br>14627   1%  96% 0.00     3632 dbuf_hash_table+0x4f8  dbuf_find<br><br><br><br></div><div>Writes are not performing htable lookups, while reads are.<br><br></div><div>-Warren V<br></div><div><br><br><br><br><br></div></div></div></div></div><div class="gmail_extra"><br><div class="gmail_quote"><div><div class="h5">On Mon, Mar 2, 2015 at 3:14 AM, Joerg Goltermann <span dir="ltr"><<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a>></span> wrote:<br></div></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div><div class="h5">Hi,<br><br>
I would try *one* TPG which includes both interface addresses<br>
and I would double check for packet drops on the Catalyst.<br><br>
The 3560 supports only receive flow control which means, that<br>
a sending 10Gbit port can easily overload a 1Gbit port.<br>
Do you have flow control enabled?<br><br>
 - Joerg<div><div><br><br>
On 02.03.2015 09:22, W Verb via illumos-developer wrote:<br></div></div></div></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div><div class="h5"><div><div>
Hello Garrett,<br><br>
No, no 802.3ad going on in this config.<br><br>
Here is a basic schematic:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQVkVqcE5OQUJyUUU/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQVkVqcE5OQUJyUUU/<u></u>view?usp=sharing</a><br><br>
Here is the Nexenta MPIO iSCSI Setup Document that I used as a guide:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQbjEyUTBjN2tTNWM/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQbjEyUTBjN2tTNWM/<u></u>view?usp=sharing</a><br><br>
Note that I am using an MTU of 3000 on both the 10G and 1G NICs. The<br>
switch is set to allow 9148-byte frames, and I'm not seeing any<br>
errors/buffer overruns on the switch.<br><br>
Here is a screenshot of a packet capture from a read operation on the<br>
guest OS (from it's local drive, which is actually a VMDK file on the<br>
storage server). In this example, only a single 1G ESXi kernel interface<br>
(vmk1) is bound to the software iSCSI initiator.<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQa2NYdXhpZkpkbU0/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQa2NYdXhpZkpkbU0/<u></u>view?usp=sharing</a><br><br>
Note that there's a nice, well-behaved window sizing process taking<br>
place. The ESXi decreases the scaled window by 11 or 12 for each ACK,<br>
then bumps it back up to 512.<br><br>
Here is a similar screenshot of a single-interface write operation:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQbU1RZHRnakxDSFU/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQbU1RZHRnakxDSFU/<u></u>view?usp=sharing</a><br><br>
There are no pauses or gaps in the transmission rate in the<br>
single-interface transfers.<br><br><br>
In the next screenshots, I have enabled an additional 1G interface on<br>
the ESXi host, and bound it to the iSCSI initiator. The new interface is<br>
bound to a separate physical port, uses a different VLAN on the switch,<br>
and talks to a different 10G port on the storage server.<br><br>
First, let's look at a write operation on the guest OS, which happily<br>
pumps data at near-line-rate to the storage server.<br><br>
Here is a sequence number trace diagram. Note how the transfer has a<br>
nice, smooth increment rate over the entire transfer.<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQWHNIa0drWnNxMmM/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQWHNIa0drWnNxMmM/<u></u>view?usp=sharing</a><br><br>
Here are screenshots from packet captures on both 1G interfaces:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQRWhyVVQ4djNaU3c/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQRWhyVVQ4djNaU3c/<u></u>view?usp=sharing</a><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQaTVjTEtTRloyR2c/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQaTVjTEtTRloyR2c/<u></u>view?usp=sharing</a><br><br>
Note how we again see nice, smooth window adjustment, and no gaps in<br>
But now, let's look at the problematic two-interface Read operation.<br>
First, the sequence graph:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQTzdFVWdQMWZ6LUU/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQTzdFVWdQMWZ6LUU/<u></u>view?usp=sharing</a><br><br>
As you can see, there are gaps and jumps in the transmission throughout<br>
the transfer.<br>
It is very illustrative to look at captures of the gaps, which are<br>
occurring on both interfaces:<br><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQc0VISXN6eVFwQzg/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQc0VISXN6eVFwQzg/<u></u>view?usp=sharing</a><br><a href="https://drive.google.com/file/d/0BwyUMjibonYQVFREUHp3TGFiUU0/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQVFREUHp3TGFiUU0/<u></u>view?usp=sharing</a><br><br>
As you can see, there are ~.4 second pauses in transmission from the<br>
storage server, which kills the transfer rate.<br>
It's clear that the ESXi box ACKs the prior iSCSI operation to<br>
completion, then makes a new LUN request, which the storage server<br>
immediately replies to. The ESXi ACKs the response packet from the<br>
storage server, then waits...and waits....and waits... until eventually<br>
the storage server starts transmitting again.<br><br>
Because the pause happens while the ESXi client is waiting for a packet<br>
from the storage server, that tells me that the gaps are not an artifact<br>
of traffic being switched between both active interfaces, but are<br>
actually indicative of short hangs occurring on the server.<br><br>
Having a pause or two in transmission is no big deal, but in my case, it<br>
is happening constantly, and dropping my overall read transfer rate down<br>
to 20-60MB/s, which is slower than the single interface transfer rate<br>
Decreasing the MTU makes the pauses shorter, increasing them makes the<br>
pauses longer.<br><br>
Another interesting thing is that if I set the multipath io interval to<br>
3 operations instead of 1, I get better throughput. In other words, the<br>
less frequently I swap IP addresses on my iSCSI requests from the ESXi<br>
unit, the fewer pauses I see.<br><br>
Basically, COMSTAR seems to choke each time an iSCSI request from a new<br>
IP arrives.<br><br>
Because the single interface transfer is near line rate, that tells me<br>
that the storage system (mpt_sas, zfs, etc) is working fine. It's only<br>
when multiple paths are attempted that iSCSI falls on its face during reads.<br><br>
All of these captures were taken without a cache device being attached<br>
to the storage zpool, so this isn't looking like some kind of ZFS ARC<br>
problem. As mentioned previously, local transfers to/from the zpool are<br>
showing ~300-500 MB/s rates over long transfers (10G+).<br><br>
-Warren V<br><br>
On Sun, Mar 1, 2015 at 9:11 PM, Garrett D'Amore <<a href="mailto:garrett@damore.org" target="_blank">garrett@damore.org</a><br></div></div><span>
<mailto:<a href="mailto:garrett@damore.org" target="_blank">garrett@damore.org</a>>> wrote:<br><br>
    I’m not sure I’ve followed properly.  You have *two* interfaces.<br>
    You are not trying to provision these in an aggr are you? As far as<br>
    I’m aware, VMware does not support 802.3ad link aggregations.  (Its<br>
    possible that you can make it work with ESXi if you give the entire<br>
    NIC to the guest — but I’m skeptical.)  The problem is that if you<br>
    try to use link aggregation, some packets (up to half!) will be<br>
    lost.  TCP and other protocols fare poorly in this situation.<br><br>
    Its possible I’ve totally misunderstood what you’re trying to do, in<br>
    which case I apologize.<br><br>
    The idle thing is a red-herring — the cpu is waiting for work to do,<br>
    probably because packets haven’t arrived (or where dropped by the<br>
    hypervisor!)  I wouldn’t read too much into that except that your<br>
    network stack is in trouble.  I’d look a bit more closely at the<br>
    kstats for tcp — I suspect you’ll see retransmits or out of order<br>
    values that are unusually high — if so this may help validate my<br>
    theory above.<br><br>
    - Garrett<br><br></span><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><span>
    On Mar 1, 2015, at 9:03 PM, W Verb via illumos-developer<br></span>
    <<a href="mailto:developer@lists.illumos.org" target="_blank">developer@lists.illumos.org</a> <mailto:<a href="mailto:developer@lists.illumos.org" target="_blank">developer@lists.<u></u>illumos.org</a>>><div><div><br>
    Hello all,<br><br><br>
    Well, I no longer blame the ixgbe driver for the problems I'm seeing.<br><br><br>
    I tried Joerg's updated driver, which didn't improve the issue. So<br>
    I went back to the drawing board and rebuilt the server from scratch.<br><br>
    What I noted is that if I have only a single 1-gig physical<br>
    interface active on the ESXi host, everything works as expected.<br>
    As soon as I enable two interfaces, I start seeing the performance<br>
    problems I've described.<br><br>
    Response pauses from the server that I see in TCPdumps are still<br>
    leading me to believe the problem is delay on the server side, so<br>
    I ran a series of kernel dtraces and produced some flamegraphs.<br><br><br>
    This was taken during a read operation with two active 10G<br>
    interfaces on the server, with a single target being shared by two<br>
    tpgs- one tpg for each 10G physical port. The host device has two<br>
    1G ports enabled, with VLANs separating the active ports into<br>
    10G/1G pairs. ESXi is set to multipath using both VLANS with a<br>
    round-robin IO interval of 1.<br><br>
    <a href="https://drive.google.com/file/d/0BwyUMjibonYQd3ZYOGh4d2pteGs/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQd3ZYOGh4d2pteGs/<u></u>view?usp=sharing</a><br><br><br>
    This was taken during a write operation:<br><br>
    <a href="https://drive.google.com/file/d/0BwyUMjibonYQMnBtU1Q2SXM2ams/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQMnBtU1Q2SXM2ams/<u></u>view?usp=sharing</a><br><br><br>
    I then rebooted the server and disabled C-State, ACPI T-State, and<br>
    general EIST (Turbo boost) functionality in the CPU.<br><br>
    I when I attempted to boot my guest VM, the iSCSI transfer<br>
    gradually ground to a halt during the boot loading process, and<br>
    the guest OS never did complete its boot process.<br><br>
    Here is a flamegraph taken while iSCSI is slowly dying:<br><br>
    <a href="https://drive.google.com/file/d/0BwyUMjibonYQM21JeFZPX3dZWTg/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQM21JeFZPX3dZWTg/<u></u>view?usp=sharing</a><br><br><br>
    I edited out cpu_idle_adaptive from the dtrace output and<br>
    regenerated the slowdown graph:<br><br>
    <a href="https://drive.google.com/file/d/0BwyUMjibonYQbTVwV3NvXzlPS1E/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQbTVwV3NvXzlPS1E/<u></u>view?usp=sharing</a><br><br><br>
    I then edited cpu_idle_adaptive out of the speedy write operation<br>
    and regenerated that graph:<br><br>
    <a href="https://drive.google.com/file/d/0BwyUMjibonYQeWFYM0pCMDZ1X2s/view?usp=sharing" target="_blank">https://drive.google.com/file/<u></u>d/<u></u>0BwyUMjibonYQeWFYM0pCMDZ1X2s/<u></u>view?usp=sharing</a><br><br><br>
    I have zero experience with interpreting flamegraphs, but the most<br>
    significant difference I see between the slow read example and the<br>
    fast write example is in unix`thread_start --> unix`idle. There's<br>
    a good chunk of "unix`i86_mwait" in the read example that is not<br>
    present in the write example at all.<br><br>
    Disabling the l2arc cache device didn't make a difference, and I<br>
    had to reenable EIST support on the CPU to get my VMs to boot.<br><br>
    I am seeing a variety of bug reports going back to 2010 regarding<br>
    excessive mwait operations, with the suggested solutions usually<br>
    being to set "cpupm enable poll-mode" in power.conf. That change<br>
    also had no effect on speed.<br><br>
    -Warren V<br><br><br><br><br>
    -----Original Message-----<br><br>
    From: Chris Siebenmann [mailto:<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a>]<br><br>
    Sent: Monday, February 23, 2015 8:30 AM<br><br>
    To: W Verb<br><br>
    Cc: <a href="mailto:omnios-discuss@lists.omniti.com" target="_blank">omnios-discuss@lists.omniti.<u></u>com</a><br></div></div>
    <mailto:<a href="mailto:omnios-discuss@lists.omniti.com" target="_blank">omnios-discuss@lists.<u></u>omniti.com</a>>; <a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a><br>
    <mailto:<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a>><span><br><br>
    Subject: Re: [OmniOS-discuss] The ixgbe driver, Lindsay Lohan, and<br>
    the Greek economy<br><br><br>
    > Chris, thanks for your specific details. I'd appreciate it if you<br><br>
    > could tell me which copper NIC you tried, as well as to pass on the<br><br>
    > iSCSI tuning parameters.<br><br><br>
    Our copper NIC experience is with onboard X540-AT2 ports on<br>
    SuperMicro hardware (which have the guaranteed 10-20 msec lock<br>
    hold) and dual-port 82599EB TN cards (which have some sort of<br>
    driver/hardware failure under load that eventually leads to<br>
    2-second lock holds). I can't recommend either with the current<br>
    driver; we had to revert to 1G networking in order to get stable<br>
    The iSCSI parameter modifications we do, across both initiators<br>
    and targets, are:<br><br><br>
    maxrecvdataseglen128k[only on Linux backends]<br><br></span>
    maxxmitdataseglen128k[only on Linux backends]<span><br><br><br>
    The OmniOS initiator doesn't need tuning for more than the first<br>
    two parameters; on the Linux backends we tune up all four. My<br>
    extended thoughts on these tuning parameters and why we touch them<br>
    can be found<br><br>
    <a href="http://utcc.utoronto.ca/~cks/space/blog/tech/UnderstandingiSCSIProtocol" target="_blank">http://utcc.utoronto.ca/~cks/<u></u>space/blog/tech/<u></u>UnderstandingiSCSIProtocol</a><br><br>
    <a href="http://utcc.utoronto.ca/~cks/space/blog/tech/LikelyISCSITuning" target="_blank">http://utcc.utoronto.ca/~cks/<u></u>space/blog/tech/<u></u>LikelyISCSITuning</a><br><br><br>
    The short version is that these parameters probably only make a<br>
    small difference but their overall goal is to do 128KB ZFS reads<br>
    and writes in single iSCSI operations (although they will be<br>
    fragmented at the TCP<br><br>
    layer) and to do iSCSI writes without a back-and-forth delay<br>
    between initiator and target (that's 'initialr2t no').<br><br><br>
    I think basically everyone should use InitialR2T set to no and in<br>
    fact that it should be the software default. These days only<br>
    unusually limited iSCSI targets should need it to be otherwise and<br>
    they can change their setting for it (initiator and target must<br>
    both agree to it being 'yes', so either can veto it).<br><br><br>
    - cks<br><br><br><br>
    On Mon, Feb 23, 2015 at 8:21 AM, Joerg Goltermann <<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a><br></span><div><div>
    <mailto:<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a>>> wrote:<br><br>
        I think your problem is caused by your link properties or your<br>
        switch settings. In general the standard ixgbe seems to perform<br>
        I had trouble after changing the default flow control settings<br>
        to "bi"<br>
        and this was my motivation to update the ixgbe driver a long<br>
        time ago.<br>
        After I have updated our systems to ixgbe 2.5.8 I never had any<br>
        problems ....<br><br>
        Make sure your switch has support for jumbo frames and you use<br>
        the same mtu on all ports, otherwise the smallest will be used.<br><br>
        What switch do you use? I can tell you nice horror stories about<br>
        different vendors....<br><br>
         - Joerg<br><br>
        On 23.02.2015 10:31, W Verb wrote:<br><br>
            Thank you Joerg,<br><br>
            I've downloaded the package and will try it tomorrow.<br><br>
            The only thing I can add at this point is that upon review<br>
            of my<br>
            testing, I may have performed my "pkg -u" between the<br>
            initial quad-gig<br>
            performance test and installing the 10G NIC. So this may<br>
            be a new<br>
            problem introduced in the latest updates.<br><br>
            Those of you who are running 10G and have not upgraded to<br>
            the latest<br>
            kernel, etc, might want to do some additional testing<br>
            before running the<br>
            -Warren V<br><br>
            On Mon, Feb 23, 2015 at 1:15 AM, Joerg Goltermann<br>
            <<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a> <mailto:<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a>><br></div></div><span>
            <mailto:<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a> <mailto:<a href="mailto:jg@osn.de" target="_blank">jg@osn.de</a>>>> wrote:<br><br>
                I remember there was a problem with the flow control<br>
            settings in the<br>
                driver, so I updated it a long time ago for our<br>
            internal servers to<br>
                Last weekend I integrated the latest changes from the<br>
            FreeBSD driver<br>
                to bring<br>
                the illumos ixgbe to 2.5.25 but I had no time to test<br>
            it, so it's<br>
                If you would like to give the latest driver a try you<br>
            can fetch the<br>
                kernel modules from<br></span>
            <a href="https://cloud.osn.de/index.____php/s/Fb4so9RsNnXA7r9" target="_blank">https://cloud.osn.de/index.___<u></u>_php/s/Fb4so9RsNnXA7r9</a><br>
            <<a href="https://cloud.osn.de/index.__php/s/Fb4so9RsNnXA7r9" target="_blank">https://cloud.osn.de/index.__<u></u>php/s/Fb4so9RsNnXA7r9</a>><span><br>
                <<a href="https://cloud.osn.de/index.__php/s/Fb4so9RsNnXA7r9" target="_blank">https://cloud.osn.de/index.__<u></u>php/s/Fb4so9RsNnXA7r9</a><br>
            <<a href="https://cloud.osn.de/index.php/s/Fb4so9RsNnXA7r9" target="_blank">https://cloud.osn.de/index.<u></u>php/s/Fb4so9RsNnXA7r9</a>>><br><br>
                Clone your boot environment, place the modules in the<br>
            new environment<br>
                and update the boot-archive of the new BE.<br><br>
                  - Joerg<br><br><br><br><br><br>
                On 23.02.2015 02:54, W Verb wrote:<br><br>
                    By the way, to those of you who have working<br>
            setups: please send me<br>
                    your pool/volume settings, interface linkprops,<br>
            and any kernel<br>
                    parameters you may have set.<br><br>
                    Warren V<br><br>
                    On Sat, Feb 21, 2015 at 7:59 AM, Schweiss, Chip<br>
                    <<a href="mailto:chip@innovates.com" target="_blank">chip@innovates.com</a> <mailto:<a href="mailto:chip@innovates.com" target="_blank">chip@innovates.com</a>><br></span>
            <mailto:<a href="mailto:chip@innovates.com" target="_blank">chip@innovates.com</a> <mailto:<a href="mailto:chip@innovates.com" target="_blank">chip@innovates.com</a>>>><div><div><br>
                        I can't say I totally agree with your performance<br>
                        assessment.   I run Intel<br>
                        X520 in all my OmniOS boxes.<br><br>
                        Here is a capture of nfssvrtop I made while<br>
            running many<br>
                        storage vMotions<br>
                        between two OmniOS boxes hosting NFS<br>
            datastores.   This is a<br>
                        10 host VMware<br>
                        cluster.  Both OmniOS boxes are dual 10G<br>
            connected with<br>
                        copper twin-ax to<br>
                        the in rack Nexus 5010.<br><br>
                        VMware does 100% sync writes, I use ZeusRAM<br>
            SSDs for log<br>
                        2014 Apr 24 08:05:51, load: 12.64, read:<br>
            17330243 KB,<br>
                        swrite: 15985    KB,<br>
                        awrite: 1875455  KB<br><br>
                        Ver     Client           NFSOPS   Reads<br>
            SWrites AWrites<br>
                        Commits   Rd_bw<br>
                        SWr_bw  AWr_bw    Rd_t   SWr_t   AWr_t<br>
             Com_t  Align%<br><br>
                        4          0       0<br>
             0       0<br>
                          0       0<br>
                        0       0       0       0       0       0       0<br><br>
                        4          0       0<br>
             0       0<br>
                          0       0<br>
                        0       0       0       0       0       0       0<br><br>
                        4          0       0<br>
             0       0<br>
                          0       0<br>
                        0       0       0       0       0       0       0<br><br>
                        4          0       0<br>
             0       0<br>
                          0       0<br>
                        0       0       0       0       0       0       0<br><br>
                        4       all                   1       0<br>
             0       0<br>
                          0       0<br>
                        0       0       0       0       0       0       0<br><br>
                        3          3       0<br>
             3       0<br>
                          0       1<br>
                        11       0    4806      48       0       0      85<br><br>
                        3          6       0<br>
             6       0<br>
                          0       3<br>
                        162       0     549     124       0       0<br>
                        3         11       0<br>
            10       0<br>
                          0       3<br>
                        27       0     776      89       0       0      67<br><br>
                        3         28       2<br>
            26       0<br>
                          0      10<br>
                        405       0    2572     198       0       0<br>
                        3       4606    4602<br>
             4       0<br>
                          0  294534<br>
                        3       0     723      49       0       0      99<br><br>
                        3       4905    4879<br>
            26       0<br>
                          0  312208<br>
                        311       0     735     271       0       0<br>
                        3       5515    5502<br>
            13       0<br>
                          0  352107<br>
                        77       0      89      87       0       0      99<br><br>
                        3      12095   12059<br>
            10       0<br>
                          0  763014<br>
                        39       0     249     147       0       0      99<br><br>
                        3        15401    6040<br>
             116    6354<br>
                        53  191605<br>
                        474  202346     192      96     144      83<br>
                        3       all <a href="tel:42574%2033086" value="+14257433086" target="_blank">42574 33086</a> <tel:42574%2033086><span><br>
            <tel:42574%20%20%2033086>     217<br>
                        6354      53 1913488<br>
                        1582  202300     348     138     153     105<br>
                        On Fri, Feb 20, 2015 at 11:46 PM, W Verb<br>
            <<a href="mailto:wverb73@gmail.com" target="_blank">wverb73@gmail.com</a> <mailto:<a href="mailto:wverb73@gmail.com" target="_blank">wverb73@gmail.com</a>><br></span>
                        <mailto:<a href="mailto:wverb73@gmail.com" target="_blank">wverb73@gmail.com</a><div><div><br>
            <mailto:<a href="mailto:wverb73@gmail.com" target="_blank">wverb73@gmail.com</a>>>> wrote:<br><br><br>
                            Hello All,<br><br>
                            Thank you for your replies.<br>
                            I tried a few things, and found the following:<br><br>
                            1: Disabling hyperthreading support in the<br>
            BIOS drops<br>
                            performance overall<br>
                            by a factor of 4.<br>
                            2: Disabling VT support also seems to have<br>
            some effect,<br>
                            although it<br>
                            appears to be minor. But this has the<br>
            amusing side<br>
                            effect of fixing the<br>
                            hangs I've been experiencing with fast<br>
            reboot. Probably<br>
                            by disabling kvm.<br>
                            3: The performance tests are a bit tricky<br>
            to quantify<br>
                            because of caching<br>
                            effects. In fact, I'm not entirely sure<br>
            what is<br>
                            happening here. It's just<br>
                            best to describe what I'm seeing:<br><br>
                            The commands I'm using to test are<br>
                            dd if=/dev/zero of=./test.dd bs=2M count=5000<br>
                            dd of=/dev/null if=./test.dd bs=2M count=5000<br>
                            The host vm is running Centos 6.6, and has<br>
            the latest<br>
                            vmtools installed.<br>
                            There is a host cache on an SSD local to<br>
            the host that<br>
                            is also in place.<br>
                            Disabling the host cache didn't<br>
            immediately have an<br>
                            effect as far as I could<br>
                            The host MTU set to 3000 on all iSCSI<br>
            interfaces for all<br>
                            Test 1: Right after reboot, with an ixgbe<br>
            MTU of 9000,<br>
                            the write test<br>
                            yields an average speed over three tests<br>
            of 137MB/s. The<br>
                            read test yields an<br>
                            average over three tests of 5MB/s.<br><br>
                            Test 2: After setting "ifconfig ixgbe0 mtu<br>
            3000", the<br>
                            write tests yield<br>
                            140MB/s, and the read tests yield 53MB/s.<br>
            It's important<br>
                            to note here that<br>
                            if I cut the read test short at only<br>
            2-3GB, I get<br>
                            results upwards of<br>
                            350MB/s, which I assume is local<br>
            cache-related distortion.<br><br>
                            Test 3: MTU of 1500. Read tests are up to<br>
            156 MB/s.<br>
                            Write tests yield<br>
                            about 142MB/s.<br>
                            Test 4: MTU of 1000: Read test at 182MB/s.<br>
                            Test 5: MTU of 900: Read test at 130 MB/s.<br>
                            Test 6: MTU of 1000: Read test at 160MB/s.<br>
            Write tests<br>
                            are now<br>
                            consistently at about 300MB/s.<br>
                            Test 7: MTU of 1200: Read test at 124MB/s.<br>
                            Test 8: MTU of 1000: Read test at 161MB/s.<br>
            Write at 261MB/s.<br><br>
                            A few final notes:<br>
                            L1ARC grabs about 10GB of RAM during the<br>
            tests, so<br>
                            there's definitely some<br>
                            read caching going on.<br>
                            The write operations are easier to observe<br>
            with iostat,<br>
                            and I'm seeing io<br>
                            rates that closely correlate with the<br>
            network write speeds.<br><br><br>
                            Chris, thanks for your specific details.<br>
            I'd appreciate<br>
                            it if you could<br>
                            tell me which copper NIC you tried, as<br>
            well as to pass<br>
                            on the iSCSI tuning<br>
                            I've ordered an Intel EXPX9502AFXSR, which<br>
            uses the<br>
                            82598 chip instead of<br>
                            the 82599 in the X520. If I get similar<br>
            results with my<br>
                            fiber transcievers,<br>
                            I'll see if I can get a hold of copper ones.<br><br>
                            But I should mention that I did indeed<br>
            look at PHY/MAC<br>
                            error rates, and<br>
                            they are nil.<br><br>
                            -Warren V<br><br>
                            On Fri, Feb 20, 2015 at 7:25 PM, Chris<br>
                            <<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a><br></div></div>
            <mailto:<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a>> <mailto:<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a><div><div><br>
            <mailto:<a href="mailto:cks@cs.toronto.edu" target="_blank">cks@cs.toronto.edu</a>>>><br><br>
                                    After installation and<br>
            configuration, I observed<br>
                                    all kinds of bad<br>
                                    in the network traffic between the<br>
            hosts and the<br>
                                    server. All of this<br>
                                    behavior is traced to the ixgbe<br>
            driver on the<br>
                                    storage server. Without<br>
                                    into the full troubleshooting<br>
            process, here are<br>
                                    my takeaways:<br><br>
                                   For what it's worth, we managed to<br>
            achieve much<br>
                                better line rates on<br>
                                copper 10G ixgbe hardware of various<br>
                                between OmniOS<br>
                                and CentOS 7 (I don't think we ever<br>
            tested OmniOS to<br>
                                OmniOS). I don't<br>
                                believe OmniOS could do TCP at full<br>
            line rate but I<br>
                                think we managed 700+<br>
                                Mbytes/sec on both transmit and<br>
            receive and we got<br>
                                basically disk-limited<br>
                                speeds with iSCSI (across multiple<br>
            disks on<br>
                                multi-disk mirrored pools,<br>
                                OmniOS iSCSI initiator, Linux iSCSI<br>
                                   I don't believe we did any specific<br>
            kernel tuning<br>
                                (and in fact some of<br>
                                our attempts to fiddle ixgbe driver<br>
            parameters blew<br>
                                up in our face).<br>
                                We did tune iSCSI connection<br>
            parameters to increase<br>
                                various buffer<br>
                                sizes so that ZFS could do even large<br>
                                operations in single iSCSI<br>
                                transactions. (More details available<br>
            if people are<br>
                                    10: At the wire level, the speed<br>
            problems are<br>
                                    clearly due to pauses in<br>
                                    response time by omnios. At 9000<br>
            byte frame<br>
                                    sizes, I see a good number<br>
                                    of duplicate ACKs and fast<br>
            retransmits during<br>
                                    read operations (when<br>
                                    omnios is transmitting). But below<br>
            about a<br>
                                    4100-byte MTU on omnios<br>
                                    (which seems to correlate to<br>
            4096-byte iSCSI<br>
                                    block transfers), the<br>
                                    transmission errors fade away and<br>
            we only see<br>
                                    the transmission pause<br>
                                   This is what really attracted my<br>
            attention. In<br>
                                our OmniOS setup, our<br>
                                specific Intel hardware had ixgbe<br>
            driver issues that<br>
                                could cause<br>
                                activity stalls during once-a-second<br>
            link heartbeat<br>
                                checks. This<br>
                                obviously had an effect at the TCP and<br>
            iSCSI layers.<br>
                                My initial message<br>
                                to illumos-developer sparked a potentially<br>
                                interesting discussion:<br><br><br></div></div>
            <a href="http://www.listbox.com/member/____archive/182179/2014/10/sort/____time_rev/page/16/entry/6:__405/__20141003125035:6357079A-__4B1D-__11E4-A39C-D534381BA44D/" target="_blank">http://www.listbox.com/member/<u></u>____archive/182179/2014/10/<u></u>sort/____time_rev/page/16/<u></u>entry/6:__405/__<u></u>20141003125035:6357079A-__<u></u>4B1D-__11E4-A39C-D534381BA44D/</a><br>
            <<a href="http://www.listbox.com/member/__archive/182179/2014/10/sort/__time_rev/page/16/entry/6:405/__20141003125035:6357079A-4B1D-__11E4-A39C-D534381BA44D/" target="_blank">http://www.listbox.com/<u></u>member/__archive/182179/2014/<u></u>10/sort/__time_rev/page/16/<u></u>entry/6:405/__20141003125035:<u></u>6357079A-4B1D-__11E4-A39C-<u></u>D534381BA44D/</a>><br><br>
            <<a href="http://www.listbox.com/__member/archive/182179/2014/10/__sort/time_rev/page/16/entry/6:__405/20141003125035:6357079A-__4B1D-11E4-A39C-D534381BA44D/" target="_blank">http://www.listbox.com/__<u></u>member/archive/182179/2014/10/<u></u>__sort/time_rev/page/16/entry/<u></u>6:__405/20141003125035:<u></u>6357079A-__4B1D-11E4-A39C-<u></u>D534381BA44D/</a><span><br>
            <<a href="http://www.listbox.com/member/archive/182179/2014/10/sort/time_rev/page/16/entry/6:405/20141003125035:6357079A-4B1D-11E4-A39C-D534381BA44D/" target="_blank">http://www.listbox.com/<u></u>member/archive/182179/2014/10/<u></u>sort/time_rev/page/16/entry/6:<u></u>405/20141003125035:6357079A-<u></u>4B1D-11E4-A39C-D534381BA44D/</a>>><br><br>
                                If you think this is a possibility in<br>
            your setup,<br>
                                I've put the DTrace<br>
                                script I used to hunt for this up on<br>
            the web:<br><br></span>
            <a href="http://www.cs.toronto.edu/~____cks/src/omnios-ixgbe/ixgbe_____delay.d" target="_blank">http://www.cs.toronto.edu/~___<u></u>_cks/src/omnios-ixgbe/ixgbe___<u></u>__delay.d</a><br>
            <<a href="http://www.cs.toronto.edu/~__cks/src/omnios-ixgbe/ixgbe___delay.d" target="_blank">http://www.cs.toronto.edu/~__<u></u>cks/src/omnios-ixgbe/ixgbe___<u></u>delay.d</a>><span><br><br>
            <<a href="http://www.cs.toronto.edu/~__cks/src/omnios-ixgbe/ixgbe___delay.d" target="_blank">http://www.cs.toronto.edu/~__<u></u>cks/src/omnios-ixgbe/ixgbe___<u></u>delay.d</a><br>
            <<a href="http://www.cs.toronto.edu/~cks/src/omnios-ixgbe/ixgbe_delay.d" target="_blank">http://www.cs.toronto.edu/~<u></u>cks/src/omnios-ixgbe/ixgbe_<u></u>delay.d</a>>><br><br>
                                This isn't the only potential source<br>
            of driver<br>
                                stalls by any means, it's<br>
                                just the one I found. You may also<br>
            want to look at<br>
                                lockstat in general,<br>
                                as information it reported is what led<br>
            us to look<br>
                                specifically at the<br>
                                ixgbe code here.<br><br>
                                (If you suspect kernel/driver issues,<br>
                                combined with kernel<br>
                                source is a really excellent resource.)<br><br>
                                          - cks<br><br><br><br><br><br></span>
                            OmniOS-discuss mailing list<br>
            <a href="mailto:OmniOS-discuss@lists.omniti" target="_blank">OmniOS-discuss@lists.omniti</a><br>
            <mailto:<a href="mailto:OmniOS-discuss@lists.omniti" target="_blank">OmniOS-discuss@lists.<u></u>omniti</a>>.____com<br>
                            <mailto:<a href="mailto:OmniOS-discuss@lists." target="_blank">OmniOS-discuss@lists.</a>_<u></u>_<a href="http://omniti.com/" target="_blank">omniti.com</a><br>
            <mailto:<a href="mailto:OmniOS-discuss@lists.omniti.com" target="_blank">OmniOS-discuss@lists.<u></u>omniti.com</a>>><br>
            <a href="http://lists.omniti.com/____mailman/listinfo/omnios-____discuss" target="_blank">http://lists.omniti.com/____<u></u>mailman/listinfo/omnios-____<u></u>discuss</a><br>
            <<a href="http://lists.omniti.com/__mailman/listinfo/omnios-__discuss" target="_blank">http://lists.omniti.com/__<u></u>mailman/listinfo/omnios-__<u></u>discuss</a>><br><br>
            <<a href="http://lists.omniti.com/__mailman/listinfo/omnios-__discuss" target="_blank">http://lists.omniti.com/__<u></u>mailman/listinfo/omnios-__<u></u>discuss</a><br>
            <<a href="http://lists.omniti.com/mailman/listinfo/omnios-discuss" target="_blank">http://lists.omniti.com/<u></u>mailman/listinfo/omnios-<u></u>discuss</a>>><br><br><br>
                    OmniOS-discuss mailing list<br>
            <a href="mailto:OmniOS-discuss@lists.omniti" target="_blank">OmniOS-discuss@lists.omniti</a><br>
            <mailto:<a href="mailto:OmniOS-discuss@lists.omniti" target="_blank">OmniOS-discuss@lists.<u></u>omniti</a>>.____com<br>
                    <mailto:<a href="mailto:OmniOS-discuss@lists." target="_blank">OmniOS-discuss@lists.</a>_<u></u>_<a href="http://omniti.com/" target="_blank">omniti.com</a><br>
            <mailto:<a href="mailto:OmniOS-discuss@lists.omniti.com" target="_blank">OmniOS-discuss@lists.<u></u>omniti.com</a>>><br>
            <a href="http://lists.omniti.com/____mailman/listinfo/omnios-____discuss" target="_blank">http://lists.omniti.com/____<u></u>mailman/listinfo/omnios-____<u></u>discuss</a><br>
            <<a href="http://lists.omniti.com/__mailman/listinfo/omnios-__discuss" target="_blank">http://lists.omniti.com/__<u></u>mailman/listinfo/omnios-__<u></u>discuss</a>><span><br><br>
            <<a href="http://lists.omniti.com/__mailman/listinfo/omnios-__discuss" target="_blank">http://lists.omniti.com/__<u></u>mailman/listinfo/omnios-__<u></u>discuss</a><br>
            <<a href="http://lists.omniti.com/mailman/listinfo/omnios-discuss" target="_blank">http://lists.omniti.com/<u></u>mailman/listinfo/omnios-<u></u>discuss</a>>><br><br><br>
                OSN Online Service Nuernberg GmbH, Bucher Str. 78,<br>
            90408 Nuernberg<br>
                Tel: <a href="tel:%2B49%20911%2039905-0" value="+49911399050" target="_blank">+49 911 39905-0</a> <tel:%2B49%20911%2039905-0><br>
            <tel:%2B49%20911%2039905-0> - Fax: +49 911<br>
                39905-55 <tel:%2B49%20911%2039905-55> -<br></span>
            <a href="http://www.osn.de/" target="_blank">http://www.osn.de</a> <<a href="http://www.osn.de/" target="_blank">http://www.osn.de/</a>><span><br>
                HRB 15022 Nuernberg, USt-Id: DE189301263, GF: Joerg<br>
        OSN Online Service Nuernberg GmbH, Bucher Str. 78, 90408 Nuernberg<br>
        Tel: <a href="tel:%2B49%20911%2039905-0" value="+49911399050" target="_blank">+49 911 39905-0</a> <tel:%2B49%20911%2039905-0> - Fax: +49<br>
        911 39905-55 <tel:%2B49%20911%2039905-55> - <a href="http://www.osn.de/" target="_blank">http://www.osn.de</a><br></span>
        <<a href="http://www.osn.de/" target="_blank">http://www.osn.de/</a>><span><br>
        HRB 15022 Nuernberg, USt-Id: DE189301263, GF: Joerg Goltermann<br><br><br></span>
    *illumos-developer* | Archives<br>
    <<a href="https://www.listbox.com/member/archive/182179/=now" target="_blank">https://www.listbox.com/<u></u>member/archive/182179/=now</a>><br>
    <<a href="https://www.listbox.com/member/archive/rss/182179/21239177-3604570e" target="_blank">https://www.listbox.com/<u></u>member/archive/rss/182179/<u></u>21239177-3604570e</a>><br>
    | Modify <<a href="https://www.listbox.com/member/?&" target="_blank">https://www.listbox.com/<u></u>member/?&</a>> Your Subscription<br>
    [Powered by Listbox] <<a href="http://www.listbox.com/" target="_blank">http://www.listbox.com/</a>><br><br></blockquote><br><br>
*illumos-developer* | Archives<br>
<<a href="https://www.listbox.com/member/archive/182179/=now" target="_blank">https://www.listbox.com/<u></u>member/archive/182179/=now</a>><br>
<<a href="https://www.listbox.com/member/archive/rss/182179/21175123-d0c8da4c" target="_blank">https://www.listbox.com/<u></u>member/archive/rss/182179/<u></u>21175123-d0c8da4c</a>> |<br>
<<a href="https://www.listbox.com/member/?&" target="_blank">https://www.listbox.com/<u></u>member/?&id_<u></u>secret=21175123-d92578cc</a>><span class=""><br>
Your Subscription       [Powered by Listbox] <<a href="http://www.listbox.com/" target="_blank">http://www.listbox.com</a>><br><br></span></blockquote><span class=""><div><div><br>
-- <br>
OSN Online Service Nuernberg GmbH, Bucher Str. 78, 90408 Nuernberg<br>
Tel: <a href="tel:%2B49%20911%2039905-0" value="+49911399050" target="_blank">+49 911 39905-0</a> - Fax: <a href="tel:%2B49%20911%2039905-55" value="+499113990555" target="_blank">+49 911 39905-55</a> - <a href="http://www.osn.de/" target="_blank">http://www.osn.de</a><br>
HRB 15022 Nuernberg, USt-Id: DE189301263, GF: Joerg Goltermann<br></div></div></span></blockquote></div><br></div></div><span class=""><div bgcolor="#ffffff" style="width:auto;margin:0;padding:5px;background-color:#fff;clear:both;border-top:1px solid #ccc"><table style="background-color:#fff" bgcolor="#ffffff" border="0" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td><font style="font-family:helvetica,sans-serif" color="#333333" size="1">
      <strong>illumos-developer</strong> | <a href="https://www.listbox.com/member/archive/182179/=now" style="text-decoration:none;color:#669933;border-bottom:1px solid #444444" title="Go to archives for illumos-developer" target="_blank">Archives</a>
<a border="0" href="https://www.listbox.com/member/archive/rss/182179/21239177-3604570e" style="text-decoration:none;color:#669933" title="RSS feed for illumos-developer" target="_blank"><img src="http://postlink.www.listbox.com/1851803/833487e62783d55fe81f119fb93ef644/21239177/8c349c6f.jpg?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2ZlZWQtaWNvbi0xMHgxMC5qcGc" border="0"></a>
 | <a href="https://www.listbox.com/member/?member_id=21239177&id_secret=21239177-2d0c9337" style="text-decoration:none;color:#669933;border-bottom:1px solid #444444" title="" target="_blank">Modify</a>
 Your Subscription</font></td><td align="right" valign="top"><a href="http://www.listbox.com/" style="border-bottom:none" target="_blank">
<img src="http://postlink.www.listbox.com/1851804/3379085af0f1cf7fc3708f04b4471ae2/21239177/8c349c6f.png?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2xpc3Rib3gtbG9nby1zbWFsbC5wbmc" title="Powered by Listbox" border="0"></a></td></tr></tbody></table></div></span></div></blockquote></div><br></div></div></blockquote></div><br></div>