1 2 3 Previous Next 39 Replies Latest reply on Jun 25, 2016 10:23 AM by Keyko

    Intel RST RAID1 performances various chipset, weird results, needs clarifications

    geremia

      Hello,

       

      I need some tech clarification about the RST behavior of recent chipset, specially in raid1.

      In this page Intel� Rapid Storage Technology (Intel� RST) — RAID 0, 1, 5, 10, Matrix RAID, RAID-Ready we can read: "The performance of a RAID 1 array is greater than that of a single drive because data can be read from multiple disks - the original and the mirror - simultaneously. Disk writes do not realize the same benefit because data must first be written to one drive, then mirrored to the other."

       

      Which is true for X58 and X79 (asus motherboards) on which i installed Corsair Performance Series Pro SSDs in raid1 ( just win7 single partition), running crystal disk benchmark on X79 gave me 990MB/s sequential read speed, while on X58 (which is sata2) gave me around 500MB/s, which is infact double the max speed of a single sata channel. So, wow, it really doubles the read speed in raid1!!

       

      Now i tested raid1 on P55, H67, H77 chipset but with standard HDD, seagate 500GB sata3 4k sectors 64MBcache, WD caviar blu 512byte sectors 16MBcache (all brand new, win7 sp1 fresh installed)

      Result is quite disappointing, sequential read speed is very floating/spiking, average floating too but got no more than 75MB/s, overall is much slower than such drives tested alone (not member of any array)(around 90MB/s).

      I tried diferent iRST drivers versions, with or without write cache enabled, no big diff. A real file transfer from/to single disk to raid1 array gives 50-60MB/s.

      I've noticed 1% cpu load, no ram leak of iastor, just the chipset gets very very busy, both in reading or writing.

       

      I thought the array need to be init and verified, i did, no diff.

       

      So, the question: is there something different in X* chipsets about iRST features/performance?

      Or, do i need special raid edition disks, like WD RE?

      How can i enjoy the "double read speed in raid1"? Do i need X79 chipset or do i need any other raid capable chipset but special raid editin disks?

       

      Thanks for any info, feel free to ask me further details, i can detail till the phy level if needed.

        • 1. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
          allan_intel

          Hi,

          I see you are benchmarking the chipset on RAID 1 , I also noticed that you are running benchmarks on X79 chipset with SSD drivers also tried standard hard disk drives with some media and classic chipset models. You cannot compare performance generated with SSD among regular hard disk drives.

          To answer your questions: X generation chipset were designed for performance systems. Also keep in mind you have an onboard RAID solution that is software/hardware combined, if you are looking for performance you should be thinking about a dedicated RAID card.

          Thanks

          Allan

          1 of 1 people found this helpful
          • 2. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
            geremia

            Hi,

            thanks for reply.

             

            Unfortunally it does not clears my doubt, i'll try to be short:

             

            - does H77 (and all others not X chipsets) support "double read speed in RAID1"? Y/N

             

            I've tested the verysame standard HDD in raid0 configuration and read/write speed is doubled without too much effort, so, why double read speed in raid1 is not working? (not working and performance worse than a single drive alone)

             

            reply1: it's enabled but working slowly because the chipset is not "designed for performance" (please elaborate a few more)

            reply2: it's not enabled at all

            reply3: ........

             

            If H and Z chipset does not support double read speed in raid1, no problem, i'll sell X chipset based motherboards to my customers, it's not a problem, just say it clearly and in bold. I would sugest also to put an asterisk on the page i linked at first.

            If double read speed in raid1 is supposed to work in H and Z chipset, we have an issue, right?

             

            Thanks again for your support.

            • 3. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
              geremia

              Hi again,

               

              I decided to go in deep with the issue, i took my test rig with H67 chipset, winxp on single drive, brand new couple of seagate 500GB 16MB cache 4k sector ST500DM003 which i configured as single drive, raid1 and raid0 for my tests. I've used HDtune sw which opens the physical device and does read tests, i set it to read in 64KB chunks.

              I hooked my sata protocol analyzer to one of the 2 seagate drive (it sits in the middle between drive and motherboad in case someone asks), here are some overall performance analysis screenshot i would like to discuss. D1 refers to Device-to-Host while H1 refers to Host-to-Device, all the values are comulative of about 10seconds of test.

              I hope the pics can be viewable.

               

              This is a single drive, as you can see there is no problem at all.

              single_64KBblock.png

               

              Here you can see 1 of the 2 raid0 drive (stipe size=64KB), speed is ok even if there are some protocol errors introduced by the host

              raid0_64KBblock.png

               

              Finally, here is he raid1 with lots more protocol errors and insane low xfer speed. But at least we can see that it takes half data from each drive as advertised, each drive goes to 35MB/s average, x2=70 which is what it comes out at the OS level.

              raid1_64KBblock.png

               

              About the raid1 split read (which should double the speed), at begin it does a sequential (inverted btw) read of small chunks without any sector skipping (sw does reads of 128 sectors but the raid driver or hw splits them in smaller chunks and invert the sequence), then lately it starts to send reads of 128sectors and skipping 128sectors (will be read from the other drive).

              ATAcmd_raid1.png

               

              So, the raid1 split read is there, but there are too many protocol errors introduced by the controller.

              I'm looking at the errors and....if i'm allowed to throw a personal first impression, i would say that the host controller, when set as raid0 or 1 (and maybe 5) starts to forget the priority rules for phy and link layer events, it prefers to timeout and overhead the whole xfer process.

              But again, this is only a first impression, i'm sure someone will be so kind to come downstair to explain what's going on.

               

              Thanks

              • 4. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                geremia

                Hi again,

                 

                it's all so weird, protocol errors are all about the drive not ready to get the FIS, so the host decides, with very variable timeout, to abort the state transition and go back to idle state.

                Host sends X_RDY and device reply with SYNC, looking at the sata specs, it seems the drive is looping in this state:

                LR2: L_RcvWaitFifo Transmit SYNCp.

                - if X_RDYp received from Phy and FIFO space available  ->   L_RcvChkRdy

                - if X_RDYp received from Phy and FIFO space not available  ->   L_RcvWaitFifo

                 

                The host controller stays in specs here, my first impression was wrong:

                "Following transmission of X_RDYP, if there is no returned R_RDYP received, no Link layer

                recovery action shall be attempted. The higher-level layers should eventually time out, and reset

                the interface"

                Here it timeouts after 4,5ms

                RAID1_preFIS_SYNC_issue2.png

                Here timeouts veryfast, case drive is not replying with any primitive at all

                RAID0_preFIS_quicktimeout.png

                Here good FIS transmission

                RAID1_goodFIS.png

                 

                So it seems Seagate has its own problems, how could the FIFO be not ready if host has been transmitting just primitives and not-primitive "fillers" ?!?!? But, i would expect the same problems when drive is alone (not part of a raid array), which does not happens, but happens when raid1 and i disconnect one disk (so reads all from one)....it's very weird.

                 

                Aside the link layer errors, there is plenty of idles, bus utilization is anyway very low, what is the host controller doing?

                • 6. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                  PeterUK

                  The RST for RAID 1 does not storage data in a way where by RAID 0 read speed can happen.

                   

                  Also with a lot of testing you might find using a cluster size of 8K when formatting the NTFS better for RAID 1 using RST.

                  • 7. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                    geremia

                    The RST for RAID 1 does not storage data in a way where by RAID 0 read speed can happen.

                     

                    Can you elaborate please?

                    In one of the picture i posted, you can see raid1 spllitted read (the verysame way as raid0 reads)

                     

                    Also with a lot of testing you might find using a cluster size of 8K when formatting the NTFS better for RAID 1 using RST.

                    I did tests at lower level to be more sure, ntfs cluster size is something upper layer, i'm quite sure the OS will use ATA read cmds of bigger size than the cluster size for big files being in consecutive sectors.

                     

                    Anyone else?

                    • 8. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                      PeterUK

                      RAID 1 is not done in strips which is why you never see that setting for strip size in RAID 1 instead data is copied from the primary to secondary and reads back only from a primary.

                      Geremia wrote:

                       

                      In one of the picture i posted, you can see raid1 spllitted read (the verysame way as raid0 reads)

                      Its not reading ahead so it will only read from one disk speed at a time not break up the data its reading from the two disks because it has no strips.

                       

                      Its not reading ahead so it will only read from one disk speed at a time not break up the data its reading from the two disks because it has no strips.

                      Had RAID 1 used strips it should write strip A to disk#1 and the same strip A to disk#2 and strip B to disk#1 and the same strip B to disk#2 their by having RAID 1 and RAID 0 read speeds by reading  strip A from disk#1 and strip B from disk#2 

                      • 9. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                        geremia

                        PeterUK ha scritto:

                         

                        RAID 1 is not done in strips which is why you never see that setting for strip size in RAID 1 instead data is copied from the primary to secondary and reads back only from a primary.

                        Sorry, but i would suggest to take a look at the pictures in deep (4th picture starting from top), you'll see only half of sectors are read from one disk, and that's a sata sniff of one disk member of a raid1 array. For sure it does not happens in writes, but it happens in reads. Since both disks contains the same data in raid1, RST reads half from a disk and half from the other to double the speed, which is stated by intel in the link in first post, which results to be true on the sata logs i did (pics above), the problem is just the RST being very slow instead of doubling the speed.

                        • 10. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                          PeterUK

                          Ok I just let Intel deal with your problem.

                          • 11. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                            Manus Freedom

                            I use a RAID10 on RST and I get the same performance problem. I am interested in the Intel reply.

                            • 12. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                              geremia

                              There will be no intel reply, cause the customer relationship is done by customer care, which queries the problem db and finds nothing.

                              Problems database is populated by upstair office, it's a one way direction, and if the bell does not ring n times, no one cares.

                               

                              It's not an issue of me alone, it's an issue for anyone that does raid1 whith shuch chipsets, unfortunally most people do not have a protocol analyzer to show they have the verysame thing happening, they can just say "it's slow" and get the reply "get a dedicated raid card" as the only valid reply.

                               

                              Many people does raid1 for security reason, most today motherboard chipset offers raid1, most of the standard hdd will fail sooner or later, it's very normal for today people that has TB of personal data to take care, it's no more exotic.

                               

                              A company like Intel is advertising double read speed in raid1, a customer is not telling "it's slow, can you analyze for me?", the customer is telling "i spent my own time in analyzing, it's slower than a single disk, drop the cofee and take a look, can you reply me it's not an issue with a valid argument?"

                              • 13. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                                t-1000

                                I would also like to know what's going on. My RAID1 array is slower than a single HDD.

                                • 14. Re: Intel RST RAID1 performances various chipset, weird results, needs clarifications
                                  huibu

                                  I would also be interested in this topic. For it surely just cannot be, that the Intel RST Raid 1 has such a terrible performance. It would be OK, if it has the same speed as a single disk, but much slower??? (BTW I have a Z77 chipset)

                                   

                                  Unfortunately the customer support from the Intel side is terrible. I also saw this on another topic, where the SW delivered with their SSD drives created faulty partition tables. But it seems Intel doesn't care at all...

                                  That's really a pity.

                                   

                                  But if there's any news on this topic, I would be very interested!

                                  1 2 3 Previous Next