Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
Besides any possible calculations - what´s the performance he can expect in real world practice under best circumstances?
------------------------------------------------------------
with kind regards ______________________________________creating IT solutions
Dipl.-Chem. Knut Kristan Weber
Senior Systems Engineer CAx Professional Services
science + computing ag Hagellocher Weg 73 phone +49 7071 94 57 473 72070 Tübingen, Germany fax +49 7071 94 57 411 http://www.science-computing.de
The trouble with doing something right the first time is that nobody appreciates how difficult it was.
What is a "sysstat 1" showing on the netapp? Has the customer run "statit" to determine if the disks are the bottleneck? How is the VIF configured? Single or multi? IP or MAC based? Is the switch configured correctly?
What are the clients? have the clients been tuned?
Need more detail....
On 3/19/07, Knut Kristan Weber k.weber@science-computing.de wrote:
Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
Besides any possible calculations - what´s the performance he can expect in real world practice under best circumstances?
with kind regards ______________________________________creating IT solutions
Dipl.-Chem. Knut Kristan Weber
Senior Systems Engineer CAx Professional Services
science + computing ag Hagellocher Weg 73 phone +49 7071 94 57 473 72070 Tübingen, Germany fax +49 7071 94 57 411 http://www.science-computing.de
The trouble with doing something right the first time is that nobody appreciates how difficult it was.
why would someone build aggregates confined to seperate shelves?? 3 aggregates of 14 physical disks, leaving only 12 after raid_dp? you might as well build traditional volumes then.
the whole point of aggregates is to spread the i/o across multiple shelves/controllers and across the most efficient use of spindles, is it not?
Knut:
One shelf of SATA disks will *never* give you performance. It's meant for density, for low-performance requirements.
Also, keep in mind that a given TCP connection/transfer will only work on one physical interface of a multi-mode VIF at a time; so 60MB/s is pretty close to the 75MB/s practical GB/e limit for a single connection. Perhaps if you tried two transfers at once the number would improve.
Glenn (the other one)
-----Original Message----- From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Knut Kristan Weber Sent: Monday, March 19, 2007 11:29 AM To: toasters@mathworks.com Subject: Throughput on 3020
Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
Besides any possible calculations - what´s the performance he can expect in real world practice under best circumstances?
------------------------------------------------------------
with kind regards ______________________________________creating IT solutions
Dipl.-Chem. Knut Kristan Weber
Senior Systems Engineer CAx Professional Services
science + computing ag Hagellocher Weg 73 phone +49 7071 94 57 473 72070 Tübingen, Germany fax +49 7071 94 57 411 http://www.science-computing.de
The trouble with doing something right the first time is that nobody appreciates how difficult it was.
SATA can give you great throughput as long as the reads\writes are sequential and very large (and very little fragmentation is present). This isn't the same as performance, as that implies latency...
As for the VIF question: the default is IP based - this means the filer will specify all outbound connections from certain IPs over certain physical interfaces. You can change this to 'rr' or round-robin which will allow traffic to be sent out of all interfaces in a round-robin fashion. However, this is only good for reads, not writes: the switch controls how data is sent _to_ the filer. Some switches will allow this type of round-robin behavior to the filer.
Overall, multithreaded applications would work best. Placing all of the volumes in their own aggregates is not a very good idea: the whole point of an aggregate is to 'aggregate' the disk I/O or bandwidth together for all the volumes to utilize. 60MB/s from 12-13 data disks isn't too bad for WAFL for some workloads - but depends on how sequential the data is and how large the blocks are.
Glenn
-----Original Message----- From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Dekhayser Sent: Monday, March 19, 2007 12:39 PM To: k.weber@science-computing.de; toasters@mathworks.com Subject: RE: Throughput on 3020
Knut:
One shelf of SATA disks will *never* give you performance. It's meant for density, for low-performance requirements.
Also, keep in mind that a given TCP connection/transfer will only work on one physical interface of a multi-mode VIF at a time; so 60MB/s is pretty close to the 75MB/s practical GB/e limit for a single connection. Perhaps if you tried two transfers at once the number would improve.
Glenn (the other one)
-----Original Message----- From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Knut Kristan Weber Sent: Monday, March 19, 2007 11:29 AM To: toasters@mathworks.com Subject: Throughput on 3020
Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
Besides any possible calculations - what´s the performance he can expect in real world practice under best circumstances?
------------------------------------------------------------
with kind regards ______________________________________creating IT solutions
Dipl.-Chem. Knut Kristan Weber
Senior Systems Engineer CAx Professional Services
science + computing ag Hagellocher Weg 73 phone +49 7071 94 57 473 72070 Tübingen, Germany fax +49 7071 94 57 411 http://www.science-computing.de
The trouble with doing something right the first time is that nobody appreciates how difficult it was.
Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
So with the Etherchannel or 802.3ad, he has a maximum throughput potential of about 200MB/s. (2Gbps). Do I understand this correctly?
Which load sharing method is the customer using on the port channel?
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
What is he doing to test? Copy files (Local disk might be bottleneck) dd? iozone? 60MB sustained is pretty good for a single GbE connection, and my initial guess is that his server is doing the best it can, but probably won't go any faster with a single NIC.
If he really wanted to tax the filer, have him use two load generators instead of one. If the etherchannel / link aggregation is setup in the typical fashion, he won't see higher throughput unless he uses two client machines.
Regards, Max
Unless doing jumbo frames (or maybe a TOE card) you are not going to see more than about 75MB/s per TCP gig connection (with jumbo frames I've seen north of 90MB/s). So, 75x4 ~300 MB/s for 4 ports. However, I don't think the head will push that much - a bit over 200MB/s for large sequential read is what it's rated at.
-----Original Message----- From: owner-toasters@mathworks.com on behalf of Max Reid Sent: Mon 3/19/2007 11:10 AM To: k.weber@science-computing.de Cc: toasters@mathworks.com Subject: Re: Throughput on 3020
Hello,
this is my first posting to this list.
A customer of us has a NetApp 3020 with 3 S-ATA shelfes each filled with 500 GB disks. He has a trunked 2*1 GBit network connection.
So with the Etherchannel or 802.3ad, he has a maximum throughput potential of about 200MB/s. (2Gbps). Do I understand this correctly?
Which load sharing method is the customer using on the port channel?
He has 3 volumes, each volume resides on a different aggregate built from the disks of one shelf each.
Via NFS he gets a perfomrance of sustained 60 MB per second with large and small files. Only one connection used during the test, no aditional traffic on the system and network segment. He expected the performance to be "much" better.
What is he doing to test? Copy files (Local disk might be bottleneck) dd? iozone? 60MB sustained is pretty good for a single GbE connection, and my initial guess is that his server is doing the best it can, but probably won't go any faster with a single NIC.
If he really wanted to tax the filer, have him use two load generators instead of one. If the etherchannel / link aggregation is setup in the typical fashion, he won't see higher throughput unless he uses two client machines.
Regards, Max
-
Glenn Dekhayser -
Glenn Walker -
Jason Herring -
Knut Kristan Weber -
Leeds, Daniel -
Max Reid -
tmac