Donner’s Daily Dose of Drama

I wanted to know exactly how much throughput I get from my new switch (HP ProcCurve 1400-24G) and found a network performance test tool that is available on many platforms - iPerf. I installed it on my Pc (Asus P5NSLI MB with dual-core Intel Pentium 4 3.4 GHz, a Raid 0 configuration, and a Marvell on-board Ethernet controller), a Synology Rackstation RS-407 NAS (ARM CPUrunning Linux, Raid 5), and a Sun Fire V20z rack server with two 2.4GHz Athlon processors running Windows Server 2003 x64 (2 Broadcom NetXtreme Gigabit Ethernet controllers, single SCSI disk).

I purchased the HP switch because it supports Jumbo frames. I tested the transfer rate between all three stations in both directions with three different MTU settings. All NICs support Jumbo frames and the value can be configured in the NIC driver settings (Windows) or in the Synology administration UI.

MTU=1500

1500 is the default (non-jumbo). I limited the MTU size for all three network cards to 1500 and ran iPerf in all 6 constellations.

iperf -c rackstation -P 1 -i 10 -p 4000
    -l 8K -f k -t 10 -d -r -L 4001

The results did not exactly blow me away. I have a mix of Cat-5 and Cat-5e cabling in the house and was concerned that I would experience bandwidth limitations. Since the servers in the test were directly connected to the switch with newly purchased Cat-6 patch cables, I expected a significantly higher transfer rate between the servers than between the workstation and either of the servers. The workstation was connected via 15 feet of Cat-5 cable, a Cat-5 wall jack, and a Cat-6 patch cable.

iPerf reported transfer rates between 150 and 250 kbps, not exactly Gigabit-like. So I turned up the MTU and hoped for the best.

MTU=4000

The next MTU size that is supported by all three network card drivers was 4000. I expected transfer rates to go up consistently.

iPerf now reported between 100 and 400 kbps, except between the workstation and the Sun server - there the performance descreased to a fraction of what it should be.

MTU=9000

9000 is the largest supported frame size. I set the MTU to 9000 on all sides and ran the same iPerf tests again.

Now iPerf reports over 500 kbps download speed from the Synology NAS, which would be nice if traffic between the Sun server and the workstation wasn’t so unrealistically slow. In addition, there now are oddities between the Sun and the Synology as well. I was able to consistenly reproduce these results.

I cannot say yet why I am observing these things. I opened a support case with HP networking and I am still waiting for an answer there. I did do some basic file transfer tests with a tool, by simply copying a large 1 GB video file between the participants, and while I did not measure such low transfer rates as iPerf reported, it definitely appears that the upload speed is consistently going down when Jumbo frames are enabled.

250,000 Mbps (bits) is a little over 30 MBps (bytes). Copying a 1GB file across the network takes about 30 seconds with MTU=1500. It does not get any faster than that, but downloads become substantially slower, e.g. > 2 minutes per GB when I set MTU=9000. The SMC Gigabit switch that I replaced and that did not have Jumbo frame support had the exact same performance. So, maybe the HP switch does not really support Jumbo frames? I am still waiting for a definitive answer from support, but if you have a Giga switch that performs better with Jumbo frames I’d like to hear about it.

Needless to say, I turned Jumbo frames off again and I am running with the basic MTU of 1500. While the iPerf numbers are not impressive the subjective performance gain from upgrading to a Gigabit switch was much more dramatic. It almost seems that iPerf does not function correctly with larger MTU sizes, too.

August 21, 2009

There is an old saying, “The jug goes to the well until it breaks”. The jug started to break as recent as last year, in March 2008. One of the large American investment firms, Bear Stearns, went belly-up in the wake of a crisis of the world’s financial markets. This crisis was caused by a lack of regulation and oversight of the US mortgage industry during the 8 years the administration of George W. Bush was in office.

When you have query that does complex calculations, maybe even has function calls in the select list

select
       costly_function_A() as result_A,
       costly_function_B() as result_B
from
       table_T

and then you need to return the result of such a calculation, but also manipulate it further and return the result of this operation as well,

select
       costly_function_A() as result_A,
       costly_function_B() as result_B
       result_A/result_B as result_ratio
from
       table_T

the result is the following dilemma:

Msg 207, Level 16, State 1, Line 4
Invalid column name 'result_A'.

It is not allowed. Invoking the function twice in the query

select
       costly_function_A() as result_A,
       costly_function_B() as result_B
       costly_function_A()/costly_function_B() as result_ratio
from
       table_T

would work, but is not an option because it nearly doubles the execution time. So what does SQL Server offer in terms of query-writing tricks that let us get around this issue?

There are 2 choices - neither is particularly elegant.

First, we can put the costly stuff into a derived table where it gets executed only once, and do the secondary calculation in an outer query:

select
       costly_T.result_A,
       costly_T.result_B
       costly_T.result_A/costly_T.result_B as result_ratio
from
(
       select
              costly_function_A() as result_A,
              costly_function_B() as result_B
       from
              table_T
)
as costly_T

Or we can use a view. If you do not already use views, don’t introduce one, but go with the derived table. If you do have views, you may already have the necessary layer of abstraction that is needed here.

I installed the Synology Rackstation last night and so far I like it a lot.
.
I installed four Seagate Barracuda ST3500630AS (500GB each) and, let me tell you, this configuration is anything but quiet. I have it sitting in a 42U rack cabinet in my basement, which is now filled with the airplane-like hum of four harddisk drives and nine fans.
Second, it is slow. The Gigabit network connection does not help if the unit’s processor is hopelessly overwhelmed - which it is even when I am the only user. I am in the process of copying content from a server with a 100MBit network connection, while the NAS is on a Gigabit connection. At the same time I am configuring Joomla, which I just installed on the Synology, and every page request takes long enough to return that I have time to write this report. Response time on any web page, photo station, Joomla, or administration page, is somewhere between 3 and 5 seconds, even when the shares are idle. It is possible that the server is still busy indexing my content, so I will give it a few days and test performance again.

< ... a few hours later ...>
So I just updated the firmware to enable SSH access to the NAS, which worked beautifully, and was able to connect right away. Luckily, top is available and I noticed that a convert process is running and routinely taking up to 100% CPU. I assume that it is still creating thumbnails for the 10’s of thousands of pictures that I uploaded earlier today. This would account for the sluggish performance.

< ... the next day ...>
I was able to install the Bootstrap and get iperf installed on the Rackstation. With both my workstation and the Rackstation connected to the new Gigabit switch, I had somewhat higher expectations regarding transfer rates than what I am actually measuring:

------------------------------------------------------------
Client connecting to 10.10.1.176, TCP port 4000
TCP window size: 102 KByte (default)
------------------------------------------------------------
[ 7] local 10.10.1.109 port 3044 connected with 10.10.1.176 port 4000
[ 6] 0.0-10.0 sec 118 MBytes 98.9 Mbits/sec
[ 6] MSS size 4034 bytes (MTU 4074 bytes, unknown interface)
[ 7] 0.0-10.0 sec 303 MBytes 255 Mbits/sec
[ 7] MSS size 4022 bytes (MTU 4062 bytes, unknown interface)


After a little bit of Googling I set the MTU size of the network card on the workstation to 9000. The results are better:

------------------------------------------------------------
Client connecting to 10.10.1.176, TCP port 4000
TCP window size: 109 KByte (default)
------------------------------------------------------------
[ 7] local 10.10.1.109 port 3441 connected with 10.10.1.176 port 4000
[ 7] 0.0-10.0 sec 575 MBytes 482 Mbits/sec
[ 7] MSS size 8948 bytes (MTU 8988 bytes, unknown interface)
[ 6] 0.0-10.6 sec 712 KBytes 549 Kbits/sec
[ 6] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)


After making sure that the patch cables on both ends are Cat 5e (is Cat 5 Enhanced the same as Cat 5e?), the numbers went up a little still, but since the cable in the wall is Cat 5 and will not be replaced unless absolutely necessary … I suppose this is as good as Gigabit Ethernet gets. I ordered Cat 6 patch cables anyways, since there will be more Gigabit traffic in the house soon. We’ll see if it makes a difference.


eth0 Link encap:Ethernet HWaddr 00:11:32:01:86:86
inet addr:10.10.1.109 Bcast:10.10.1.255 Mask:255.255.255.0
UP BROADCAST RUNNING MULTICAST MTU:9004 Metric:1
RX packets:63186085 errors:0 dropped:0 overruns:0 frame:0
TX packets:37736653 errors:0 dropped:0 overruns:0 carrier:0
collisions:0 txqueuelen:512
RX bytes:2705338793 (2.5 GiB) TX bytes:3908546493 (3.6 GiB)
Interrupt:21


My switch is the SMC GS16 (no jumbo frame support, or I am sure SMC would mention it on their web site). Why I get a 500% improvement in throughput by enabling Jumbo frames on the client is beyond me. I was hoping that by plugging my laptop directly into the switch, with the RS407 plugged directly into the switch, I would get better transfer rates again, but no. It appears that the Intel PRO 1000 network controller in my Lenovo T60p laptop does not support Jumbo frames, at least there is no way of enabling them in the most recent drivers. The desktop is an Asus P5NSLI and enabling Jumbo frames on the onboard Marvell Yukon PCI Gigabit Ethernet Controller really seems to make all the difference.

The HTTP performance is still poor. The image thumbnails are apparently all created, since the machine is mostly idling along (why it does not go into hibernation is yet another question to answer), but it still responds slower to page requests than most external sites that I go to. Apache uses almost 100% of CPU time when it puts the HTTP response together - and these are light-weight pages.

Overall, I still like the Synology. It is comparatively expensive - the Acer Aspire Easystore withg 2 TB (the device is not marketed in the US) is available for almost half of what I paid. Still, a read transfer rate of almost 30MB/s over the network is pretty cool. Copying an 800MB VCD image from the NAS to the local disk takes less than 30s. I am looking forward to filling up all this space now.

< ... after one week ...>

I just noticed that the iPerf results above were Kilobits, not Megabits. Now it starts to make sense. Setting the MTU to 9 kB does not work after all if the switch does not support it. Or whatever else causes the slowdown. Since I did not really notice a problem with transfers, I am not so sure how meaningful these results are anyways. I set the MTU back to 1.5 kB. The Cat 6 patch cables arrived, btw, and I did not notice a difference.

Useful resources:

• Instructions for how to install the Bootstrap
• ARM/Marvell Bootstrap (required for ipkg and pretty much any customization work)
• More about my experiments with the Rackstation’s network performance

< ... update ...>
Synology released new firmware for the Rackstation a few months back. I have been using it for some time. The new Web interface really makes the product much more usable. There is now a Surveillance station that captures camera streams and records to disk. Unfortunately, it does not work with my camera. Still, if you don’t have it, I highly recommend getting the update.