Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[InflatableMouse]

I must have tempted the virtual network spirits because I'm now having performance issues. Some very knowledgable folks here so hopefully you can help me resolve this issue I'm having.

I have 2 HP ProCurve 1810 switches (called FRONT and BACK), connected to each other via Port 1. My pc is connected to FRONT and the server (HTPC) connects to BACK. Simple. Router is hooked up to port 7 on FRONT.

When I bought the switches they were running firmware 1.17. I upgraded the main firmware (they have dual firmwares) to v2.2. I'm pretty sure this is when my performance dropped but unfortunately I failed to a pre/post tests. I just did it and few days later I noticed the trouble.

I know for a fact that my copy jobs of large files averaged around 100MB/s if I left it alone (no other traffic). Typically when I was using my pc's or I had a mix of files I'd average around 80-90MB/s.

Today I only reach about 30% of that. I've flashed the backup firmware of the switches back to 1.17, I've done factory resets on both of them and I've reinstalled both pc's and downloaded the latest drivers. I've tested several drives for the NIC's actually (both use Realtek PCIe GBe integrated).

No matter what I try, I can't get it above 30-35MB/s. If I enable jumbo frames, I'm getting WAIT packages in my switches' logs. I can see it bursts, pauses, bursts, pauses. In the end, 34MB/s. Flow control doesn't seem to do anything noticable.

I have no idea what's going on ...  there is no packet loss, no errors, mismatches ... ? _{(and for the wiseguys here, yes I've tried setting gigabit fixed to 1000/full ).}

Thanks in advance.

[InflatableMouse]

Oke I *think* it's solved, need to do a few more tests because it's still not up to par with what I thought it was.

I modified the TCPAckFrequency parameter in the registry, this boosted network throughput by about 70%.

After changing that jumbo frames worked correctly too (no more pauses during copies) and it added a few percent to the speed.

During my quest to find a solution I ran into TCP Optimizer. This program may have royally screwed up my User Registry to the point Windows wouldn't even load the profile any more (the other PC had no problem so it might have been coincidence). Regardless, after I fixed that (cough) issue, it did add another few percent to local network speed and the program seems to work fine with the new profile.

I'm now averaging around ~70MB/s. iperf is able to saturate 1gbit of bandwidth too. Copy jobs start around 125MB/s with ~95% network utilization and gradually drops to 70MB/s in the end for a 4GB file.

[Sparks67]

I just went to a new Cisco SG-200 Smart switch, and I average around 125MB/sec on large copies.  This is SATA 2 drives, but I was really surprised on the performance of a new Sata 3 drive that I bought.  It peaked around 190MB/sec on large backup copy, but it was averaging around 150MB/sec on the file copy.   I almost want to ditch my Sata 2 drives and replace them with Sata 3.   

[InflatableMouse]

Averaging above 125MB/s is impossible on gigabit, the maximum bandwidth you have is 125MB/s (1000 / , and that includes overhead from TCP headers and whatnot so what's actually left for you to use is less than that.

I can see speeds of 450MB/s but that's cached and not actually real network speeds. I have never seen a situation where a speed of over 100MB/s is attained over a longer duration, like copying several gigabytes.

I think you're either confusing Mb's with MB's (megabits vs megabytes), or your using a program that's compressing data (like your backup program?).

[glynor]

I'm now averaging around ~70MB/s. iperf is able to saturate 1gbit of bandwidth too. Copy jobs start around 125MB/s with ~95% network utilization and gradually drops to 70MB/s in the end for a 4GB file.

That dropoff could just be disk performance for a long, sustained write.  Are you copying to/from a fast (SSD or RAID5) volume?

That's roughly what I see when I copy to/from a single spinning disk on my network.  If I go to/from SSDs (or from a SSD to my RAID) then it stays pretty much up around the ~118MB/s mark.  I've never bothered with Jumbo Frames, even though my switches and network cards all support it fine (though I'm not sure how that would impact my WiFi network and devices).

[mercedesman]

You haven't mentioned if you checked your negotiated connections between the switches and computers. If both sides are set to auto negotiate, there may be a mismatch which will result in a half-duplex connection and loss of bandwidth.

Network connections are measured in bits, not bytes as indicated in a previous post.

HTH

[Sparks67]

At the time, I was copying from Areca Raid 5 array to Areca Raid 6 array.  The Areca requires you to use an enterprise drive, but the speed stays rather constant.  That speed was with Sata 2 drives.  Sorry, I mis-type the post on Sata 3.  The Sata 3 was installed into the motherboard SATA 3 port.   Just doing a comparison between SATA 2 vs SATA 3.  In the past, I had a Netgear FS-108 unmanaged switch (slow), but just switching to the Cisco SG-200 (business) Smart switch,  increased my network speed for copies.  Smart switch gives you the speed where you need it.  Very different than unmanaged switch.  I have tried the Jumbo frames, but it is doesn't improve performance.  Switching to SATA 3 drives or SAS drives, will be expensive for me.  Typically, I transfer around 1TB at a time or more.  Well, I had to recently backup my 18TB RAID Array.   

The new setup that I recently built has Chenbro http://usa.chenbro.com/corporatesite/products_detail.php?sku=104 case, so should be more reliable than the junk Norco.  I replaced the fans in this case and it is rather quiet.   I just about to finish the installing the silent htpc in my rack, well, it is almost silent other than cpu fan.  I let you know on the speed from Areca RAID to SSD drives over the network.  I plan to switch my CAT6 cable for CAT6A.

[InflatableMouse]

That dropoff could just be disk performance for a long, sustained write.  Are you copying to/from a fast (SSD or RAID5) volume?

That's roughly what I see when I copy to/from a single spinning disk on my network.  If I go to/from SSDs (or from a SSD to my RAID) then it stays pretty much up around the ~118MB/s mark.  I've never bothered with Jumbo Frames, even though my switches and network cards all support it fine (though I'm not sure how that would impact my WiFi network and devices).

Now that I think of it .. I didn't realize this at first but ... both pc's have Drive Bender (which, for the sake of ease I call file striping), one is 2 disks and the other has 2 pools of 3 and 2 disks. This isn't real striping as RAID does so 1 file always goes to 1 disk. Many small files are killing for general network performance but DB actually gets faster as it can access all the disks. Local copies are slightly faster on the long run (on large files) but it really screams on copying many files.

I'll test with the SSD's, both PC's have a Samsung 830 series so that should be similar to what you're getting.

[InflatableMouse]

You haven't mentioned if you checked your negotiated connections between the switches and computers. If both sides are set to auto negotiate, there may be a mismatch which will result in a half-duplex connection and loss of bandwidth.

Network connections are measured in bits, not bytes as indicated in a previous post.

HTH

Yeh because I linked to that other thread in my OP. I always set everything fixed, I explained that there.