Re: Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
Mike Christie micha...@cs.wisc.edu schrieb am 27.08.2014 um 23:49 in Nachricht 53fe5276.2060...@cs.wisc.edu: On 08/27/2014 02:24 AM, Ulrich Windl wrote: Learner Study learner.st...@gmail.com schrieb am 27.08.2014 um 02:13 in Nachricht CAP8+hKW=HApS+=vxeaaibtbbd7yzndu4squt+84se99aglc...@mail.gmail.com: Hi Mike, Thanks for suggestions I think you meant, echo 1 /sys/block/sdX/device/delete I don't see /sys/block/sdX/device/remove in my setup. I'm not sure: Is it echo offline /sys/block/sdX/device/state, echo scsi remove-single-device ${host} ${channel} ${id} ${lun} /proc/scsi/scsi, or echo 1 /sys/class/scsi_device/${host}:${channel}:${id}:${lun}/device/delete ? To delete a device just do echo 1 /sys/block/sdX/device/delete I think the confusing thing is that you don't see a delete in /sys/block/sdX/device. You can also do it through proc if it is enabled for your kernel. No need to offline the device before deleting. The scsi layer will handle the device state transitions. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
On 08/28/2014 12:59 AM, Ulrich Windl wrote: To delete a device just do echo 1 /sys/block/sdX/device/delete I think the confusing thing is that you don't see a delete in /sys/block/sdX/device. Not sure what you mean. I do: ls /sys/block/sda/device/ block evt_media_change max_sectors rescanstate bsg generic modalias rev subsystem delete iocounterbits modelscsi_device timeout device_blocked iodone_cntpowerscsi_disk type dh_stateioerr_cnt queue_depth scsi_generic uevent driver iorequest_cnt queue_type scsi_levelvendor -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
On 08/28/2014 11:29 AM, Mike Christie wrote: On 08/28/2014 12:59 AM, Ulrich Windl wrote: To delete a device just do echo 1 /sys/block/sdX/device/delete I think the confusing thing is that you don't see a delete in /sys/block/sdX/device. Not sure what you mean. I do: ls /sys/block/sda/device/ block evt_media_change max_sectors rescanstate bsg generic modalias rev subsystem delete iocounterbits modelscsi_device timeout device_blocked iodone_cntpowerscsi_disk type dh_stateioerr_cnt queue_depth scsi_generic uevent driver iorequest_cnt queue_type scsi_levelvendor Ah, I see. I think depending on the kernel config options used the device symlink might not even be there. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
Learner Study learner.st...@gmail.com schrieb am 27.08.2014 um 02:13 in Nachricht CAP8+hKW=HApS+=vxeaaibtbbd7yzndu4squt+84se99aglc...@mail.gmail.com: Hi Mike, Thanks for suggestions I think you meant, echo 1 /sys/block/sdX/device/delete I don't see /sys/block/sdX/device/remove in my setup. I'm not sure: Is it echo offline /sys/block/sdX/device/state, echo scsi remove-single-device ${host} ${channel} ${id} ${lun} /proc/scsi/scsi, or echo 1 /sys/class/scsi_device/${host}:${channel}:${id}:${lun}/device/delete? How do following FIO options look? [default] rw=read size=4g bs=1m ioengine=libaio direct=1 numjobs=1 filename=/dev/sda runtime=360 iodepth=256 Thanks for your time! On Tue, Aug 26, 2014 at 4:49 PM, Michael Christie micha...@cs.wisc.edu wrote: On Aug 26, 2014, at 3:11 PM, Learner learner.st...@gmail.com wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? There is no iscsi initiator or target setting for this. The default is to show all paths (each /dev/sdx is a path to the same device).. You would have to manually delete some paths by doing echo 1 /sys/block/sdX/device/remove When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? How are you matching sessions with devices? It should just be a matter of running fio on the right devices. If you run: iscsiadm -m session -P 3 you can see how the sdXs match up with sessions/connections. If you run fio to a /dev/sdX from each session, you should be seeing IO to all 3 sessions. Thanks! Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
On Tue, 26 Aug 2014 13:05:11 -0700 Learner learner.st...@gmail.com wrote: How many iscsi and underlying top sessions are u using? If multiple, pls check if all to sessions are being used. Btw, what tuning did u perform to fix Tcp BDP issue? I'm just doing netcat tests to/from /dev/shm at the moment. I wouldn't consider it fixed necessarily, but the info from this link was useful: http://www.kehlet.cx/articles/99.html Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
I had applied the tuning for my 10g link but didn't see much impact. Actually for me tcp is already line rate with 2/3 threads but iscsi/fio read is around 5.5gbps only - with 3/4 fio threads. Perhaps the bottleneck is somewhere else... Thanks! Sent from my iPhone On Aug 27, 2014, at 8:25 AM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 13:05:11 -0700 Learner learner.st...@gmail.com wrote: How many iscsi and underlying top sessions are u using? If multiple, pls check if all to sessions are being used. Btw, what tuning did u perform to fix Tcp BDP issue? I'm just doing netcat tests to/from /dev/shm at the moment. I wouldn't consider it fixed necessarily, but the info from this link was useful: http://www.kehlet.cx/articles/99.html Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
On 08/27/2014 02:24 AM, Ulrich Windl wrote: Learner Study learner.st...@gmail.com schrieb am 27.08.2014 um 02:13 in Nachricht CAP8+hKW=HApS+=vxeaaibtbbd7yzndu4squt+84se99aglc...@mail.gmail.com: Hi Mike, Thanks for suggestions I think you meant, echo 1 /sys/block/sdX/device/delete I don't see /sys/block/sdX/device/remove in my setup. I'm not sure: Is it echo offline /sys/block/sdX/device/state, echo scsi remove-single-device ${host} ${channel} ${id} ${lun} /proc/scsi/scsi, or echo 1 /sys/class/scsi_device/${host}:${channel}:${id}:${lun}/device/delete? To delete a device just do echo 1 /sys/block/sdX/device/delete You can also do it through proc if it is enabled for your kernel. No need to offline the device before deleting. The scsi layer will handle the device state transitions. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Antw: Re: Best way to create multiple TCP flows on 10 Gbps link
Mark Lehrer m...@knm.org schrieb am 25.08.2014 um 20:58 in Nachricht ximss-10382...@knm.org: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit setup, I seem to Hi! You are referring to networks sockets, not to CPU sockets, I guess. Have you tried larger packets (if you can control the LAN). I don't know if open iSCSI can do IPv6, but from what I read IPv6 could give better TCP performance. Have you checked interrupt assignments for the NIC? I guess your card is PCIe and it uses one lane? Have you tried (for comparison) to do just a netcat to/from /dev/zero? You have to analyze the groups, hardware, network stack and iSCSI separately, I guess. iSCSI can not do any better than the networks stack, and the network stack cannot do better than the hardware can. Regards, Ulrich hit a wall around 3 gigabits when doing a single TCP socket. To go far above that I need to do multipath, initiator-side RAID, or RDMA. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html On 08/25/2014 02:58 PM, Mark Lehrer wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit setup, I seem to hit a wall around 3 gigabits when doing a single TCP socket. To go far above that I need to do multipath, initiator-side RAID, or RDMA. Thanks, Mark -- Alvin Starr. -- Alvin Starr || voice: (416)585-9971x690 Interlink Connectivity|| fax: (416)585-9974 al...@iplink.net || -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
iperf performance for TCP is line rate in both directions using 3 threads However, I can just get 700MB/s Write and 570MB/s Reads with iSCSI. Thanks for any pointers! On Tuesday, August 26, 2014 1:11:59 PM UTC-7, learner.study wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? Thanks! Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. On Tuesday, August 26, 2014 1:11:59 PM UTC-7, learner.study wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? Thanks! Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
I have a couple of iscsi links running on 1G and not in your range of hw and demand at all. I ran an ISP for about 20 years and got bitten by the BDP a number of times now so when someone describes the problem I know what to look for. On 08/26/2014 04:05 PM, Learner wrote: How many iscsi and underlying top sessions are u using? If multiple, pls check if all to sessions are being used. Btw, what tuning did u perform to fix Tcp BDP issue? Thanks Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- Alvin Starr. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
On Aug 26, 2014, at 3:11 PM, Learner learner.st...@gmail.com wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? There is no iscsi initiator or target setting for this. The default is to show all paths (each /dev/sdx is a path to the same device).. You would have to manually delete some paths by doing echo 1 /sys/block/sdX/device/remove When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? How are you matching sessions with devices? It should just be a matter of running fio on the right devices. If you run: iscsiadm -m session -P 3 you can see how the sdXs match up with sessions/connections. If you run fio to a /dev/sdX from each session, you should be seeing IO to all 3 sessions. Thanks! Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
Hi Mike, Thanks for suggestions I think you meant, echo 1 /sys/block/sdX/device/delete I don't see /sys/block/sdX/device/remove in my setup. How do following FIO options look? [default] rw=read size=4g bs=1m ioengine=libaio direct=1 numjobs=1 filename=/dev/sda runtime=360 iodepth=256 Thanks for your time! On Tue, Aug 26, 2014 at 4:49 PM, Michael Christie micha...@cs.wisc.edu wrote: On Aug 26, 2014, at 3:11 PM, Learner learner.st...@gmail.com wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? There is no iscsi initiator or target setting for this. The default is to show all paths (each /dev/sdx is a path to the same device).. You would have to manually delete some paths by doing echo 1 /sys/block/sdX/device/remove When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? How are you matching sessions with devices? It should just be a matter of running fio on the right devices. If you run: iscsiadm -m session -P 3 you can see how the sdXs match up with sessions/connections. If you run fio to a /dev/sdX from each session, you should be seeing IO to all 3 sessions. Thanks! Sent from my iPhone On Aug 26, 2014, at 12:53 PM, Mark Lehrer m...@knm.org wrote: On Tue, 26 Aug 2014 08:58:46 -0400 Alvin Starr al...@iplink.net wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit You are likely getting hit by the bandwidth-delay product. Take a look at http://en.wikipedia.org/wiki/Bandwidth-delay_product and http://www.kehlet.cx/articles/99.html Thanks that helped get my netcat transfer up over 500MB/sec using IPoIB. Unfortunately that is still only about 10% of the available bandwidth. I'll keep on tweaking and see how far I can take it. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
On Aug 26, 2014, at 6:49 PM, Michael Christie micha...@cs.wisc.edu wrote: On Aug 26, 2014, at 3:11 PM, Learner learner.st...@gmail.com wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? There is no iscsi initiator or target setting for this. The default is to show all paths (each /dev/sdx is a path to the same device).. You would have to manually delete some paths by doing echo 1 /sys/block/sdX/device/remove When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? How are you matching sessions with devices? It should just be a matter of running fio on the right devices. If you run: iscsiadm -m session -P 3 you can see how the sdXs match up with sessions/connections. If you run fio to a /dev/sdX from each session, you should be seeing IO to all 3 sessions. How are you determining if a session is being used or not? Are you running the iscsiadm -m session --stats command, watching with wireshark/tcpdump or something else? If you have all three IPs on the same subnet, then it is going to be a little more complicated than what I described above. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
I am monitoring with netstat -a...looking at sendq and recvq there for the three iscsi/tcp sessions. Also checked with tcpdump. Thanks! Sent from my iPhone On Aug 26, 2014, at 9:46 PM, Michael Christie micha...@cs.wisc.edu wrote: On Aug 26, 2014, at 6:49 PM, Michael Christie micha...@cs.wisc.edu wrote: On Aug 26, 2014, at 3:11 PM, Learner learner.st...@gmail.com wrote: Another related observation and some questions; I am using open iscsi on init with IET on trgt over a single 10gbps link There are three ip aliases on each side I have 3 ramdisks exported by IET to init I do iscsi login 3 times, once using each underlying ip address and notice that each iscsi session sees all 3 disks. Is it possible to restrict such that each init only sees one separate disk? There is no iscsi initiator or target setting for this. The default is to show all paths (each /dev/sdx is a path to the same device).. You would have to manually delete some paths by doing echo 1 /sys/block/sdX/device/remove When I run fio on each mounted disk, I see that only two underlying tcp sessions are being used - that limits the perf. Any ideas on how to overcome this? How are you matching sessions with devices? It should just be a matter of running fio on the right devices. If you run: iscsiadm -m session -P 3 you can see how the sdXs match up with sessions/connections. If you run fio to a /dev/sdX from each session, you should be seeing IO to all 3 sessions. How are you determining if a session is being used or not? Are you running the iscsiadm -m session --stats command, watching with wireshark/tcpdump or something else? If you have all three IPs on the same subnet, then it is going to be a little more complicated than what I described above. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
I find upping some of the default Linux network params helps with throughput Edit /etc/sysctl.conf, then update the system using #sysctl –p # Increase network buffer sizes net.core.rmem_max = 16777216 net.core.wmem_max = 16777216 net.ipv4.tcp_rmem = 8192 87380 16777216 net.ipv4.tcp_wmem = 4096 65536 16777216 net.core.wmem_default = 262144 net.core.rmem_default = 262144 net.core.wmem_default = 262144 net.core.rmem_default = 262144 I also find that increasing the disk read ahead really helps with sequential read loads. blockdev –setra X device name i.e. #blockdev –setra 4096 /dev/sda or /dev/mapper/mpath1 Also some small tweaks to iscsid.conf can yield some improvements. #/etc/iscsi/iscsid.conf node.session.cmds_max = 1024 --- Default is 128 node.session.queue_depth = 128 --- Default is node.conn[0].iscsi.MaxRecvDataSegmentLength = 131072 --- try 64K-512K On Mon, Aug 25, 2014 at 2:58 PM, Mark Lehrer m...@knm.org wrote: I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? On my 40 gigabit setup, I seem to hit a wall around 3 gigabits when doing a single TCP socket. To go far above that I need to do multipath, initiator-side RAID, or RDMA. Thanks, Mark -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
On 08/25/2014 04:40 PM, Mark Lehrer wrote: On Mon, 25 Aug 2014 15:48:02 -0500 Mike Christie micha...@cs.wisc.edu wrote: On 08/25/2014 03:31 PM, Donald Williams wrote: On a semi-related note, are there any good guides out there to tuning Linux for maximum single-socket performance? What kernel are you using? Are you doing IO to one LU or multiple? Single socket from one machine to another; my current test platform is using either 3.13 or 3.15. I guess the main question I'm trying to answer is: is it reasonable to expect to get 2GB/sec over a single TCP socket, or should I start Were you using 10 Gb or 40 Gb? 10 right? You meant 2 Gigabit/Gb above then right? If so, then no. I can get around 5 or 6 Gb for writes using the defaults and scst and ramdisks for the LUs. The only think I have had to do in some recent kernels is turn tcp_autocorking off. I am working on a patch for that now. How are you running fio? -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
Thanks Mike - That helped On Saturday, August 23, 2014 2:41:01 AM UTC+5:30, Mike Christie wrote: On Aug 22, 2014, at 12:07 PM, Redwood Hyd redwo...@gmail.com javascript: wrote: Hi All, I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) I exported 3 Ramdisks, via 3 different IP aliases to initator, did three iscsi logins , 3 mounts points and then 3 fio jobs in parallel (256K block size each). Question 1) Is above a real use case where from same iscsi initiator i did 3 iscsi logins to same target (via different IP addresses) ? Anything pros/cons with this. This seems normal. Question 2) What are the other best ways to create parallel TCP flows (because it seems open-iscsi does'nt have MC/S support) Multiple sessions to different portals then use dm-multipath over all those paths/sessions to the LU. Question 3) In this scenario can I use dm-multipath - can someone suggest most common way so that at TCP level i get multiple flows. What you described above, when you run /sbin/multipath /sbin/multipath -ll Do you see each device having 3 paths? Did you set it up to do round robin for dm multipath path selection? If so, each path is going to be a different tcp socket connection which the iscsi initiator and dm-multipath will use to send IO on. At my last job, fusion-io/sandisk, we sold a high performance target, and to get the highest throughput when using linux we had to create extra sessions/connections to avoid some bottlenecks in the linux block/scsi layer. Above you would have a session to each target portal/ip. We would set node.session.nr_sessions in iscsid.conf to greater than one so each portal would have nr_sessions sessions/connections. When you run iscsiadm -m node -T target -p ip -l, iscsiadm would then create nr_session to that portal. iscsiadm -m session would show the extra sessions when logged in and multipath -ll should show the extra paths. You can also just do iscsiadm -m session -R SID -o new to dynamically add another session/connection. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.
Re: Best way to create multiple TCP flows on 10 Gbps link
On Aug 22, 2014, at 12:07 PM, Redwood Hyd redwood...@gmail.com wrote: Hi All, I am trying to achieve10Gbps in my single initiator/single target env. (open-iscsi and IET) I exported 3 Ramdisks, via 3 different IP aliases to initator, did three iscsi logins , 3 mounts points and then 3 fio jobs in parallel (256K block size each). Question 1) Is above a real use case where from same iscsi initiator i did 3 iscsi logins to same target (via different IP addresses) ? Anything pros/cons with this. This seems normal. Question 2) What are the other best ways to create parallel TCP flows (because it seems open-iscsi does'nt have MC/S support) Multiple sessions to different portals then use dm-multipath over all those paths/sessions to the LU. Question 3) In this scenario can I use dm-multipath - can someone suggest most common way so that at TCP level i get multiple flows. What you described above, when you run /sbin/multipath /sbin/multipath -ll Do you see each device having 3 paths? Did you set it up to do round robin for dm multipath path selection? If so, each path is going to be a different tcp socket connection which the iscsi initiator and dm-multipath will use to send IO on. At my last job, fusion-io/sandisk, we sold a high performance target, and to get the highest throughput when using linux we had to create extra sessions/connections to avoid some bottlenecks in the linux block/scsi layer. Above you would have a session to each target portal/ip. We would set node.session.nr_sessions in iscsid.conf to greater than one so each portal would have nr_sessions sessions/connections. When you run iscsiadm -m node -T target -p ip -l, iscsiadm would then create nr_session to that portal. iscsiadm -m session would show the extra sessions when logged in and multipath -ll should show the extra paths. You can also just do iscsiadm -m session -R SID -o new to dynamically add another session/connection. -- You received this message because you are subscribed to the Google Groups open-iscsi group. To unsubscribe from this group and stop receiving emails from it, send an email to open-iscsi+unsubscr...@googlegroups.com. To post to this group, send email to open-iscsi@googlegroups.com. Visit this group at http://groups.google.com/group/open-iscsi. For more options, visit https://groups.google.com/d/optout.