SQL Centre of Excellence

With a “ragged” hierarchy we often need a way to make sure that the hierarchy is displayed in a user friendly fashion when measures can be linked to any level (including parents).

There are two established solutions

1) The Pure DAX approach

Alberto was one of the first to publish a DAX workaround for the problem of ragged hierarchies here:

http://sqlblog.com/blogs/alberto_ferrari/archive/2011/07/19/parent-child-hierarchies-in-tabular-with-denali.aspx

This workaround starts to get very complex when you have multiple hierarchies and performance can suffer on large hierarchies as the evaluation is at run time, but it does work and is “supported”.

2) HideMemberIf

BIDS helper exposes an “unsupported” technique to add HideMemberIf into the XMLA of the model definition. https://bidshelper.codeplex.com/ 

This solution looks more appealing as a lot of the work is done during ProcessRecalc when it does Hierarchy Processing – hopefully giving benefit at query time. It also involves less calculated measures, columns and code which sounds nice.

However two words of warning:

a) There is no guarantee that this will work in later editions

b) There are stability issues with a large number of levels. On the model we are working the ProcessRecalc and visual studio will just hang when we try to set this property on a NINE level hierarchy.

 

In the end we did use HideMemberIf, but only set the property on specific levels which we know can have measures attached to them to avoid the stability issues of having this turned on for all levels.

I’ve been diving deep into how long Analysis Services can take to open a connection for a few months now and thanks to some tips from many people (Alberto Ferrari, Greg Galloway, Darren Gospel to name a few) here are some ideas to improve performance.

1. Make sure your network is not slow

OK, sounds kinda obvious, but this was killing me for months.

Most of me problems ended up being driver related with various incompatibilities between our NICs, Hyper-V and some of the new features.

A simple internet speed test or benchmarking tool would have shown that the VMs running SSAS were running very slow with packet loss. I blogged more details here

http://blogs.prodata.ie/post/Solved-slow-SSAS-connections-on-Hyper-V.aspx

2. Avoid Named Instances

It can take up to double the time to open a connection to a named instance. As a test I opened 1,000 connections to a remote server with five scenarios:

  • By Name (Default Instance)
  • By IP (Default Instance)
  • By Server Name and Instance Name (Named Instance)
  • By Name and Port number (Named Instance)
  • By IP and Port number (Named Instance)

image

The results show that if we are using a named instance (e.g. bob-pc\tabular) then connections can be almost double the time to make. Once we manually specify the port number and bypass the “browser” service, connection time is about the same for the other test cases.

There was a very small improvement for using the IP instead of the name, but this was so small I’m not sure it would be worth the effort (2.5%)

lesson learned: Use fixed port numbers when using SSAS instances. If you do need named instances and want to simplify connections it should be possible to give each instance  the same default port number (2383) and bind a different IP address to each instance. We've been using this trick on SQL DBEngine clusters for years to simplify connections and manageability.

2. Optimising Local Connections

If the client and SSAS are running on the same server (e.g. development laptop). You can optimise the connection even further by using the IP address and/or avoiding using the alias “localhost”

image

Ok these are small numbers, but its a 45% gain to be made on connection time Smile

one eye opener or me is there was a measurable difference to show that using “localhost” was not the best option for me.

3. IP4 or IP6 ?

I ran a quick test comparing use of IP4 to IP6 for connecting to an SSAS Server. This did show a 30% improvement with the older IP4 (I have no idea why this would be so!)

image

 

4. Do I care about connection time ?

Many tools like Excel either keep a connection open to SSAS or re-use a single connection, so connection time might not be that important.

Other tools may open a connection per request or in some cases (reporting services) many many connections per user – In this case the impact of faster connections may be more noticeable.

Over the last few months we have been developing reporting services reports on top of a SSAS tabular model. One of the banes of my life has been the unusable performance of our reports, especially if they open a lot of connections.

This week I finally got some time to triage the issue and found the root cause, so I’ll post it here in case anyone else is hit by this.

How Can we measure Connection Time on Reports ?

Reporting services 2012 onwards has additional performance data on datasets on the ExecutionLog3 view  in the AdditionalInfo XML column. In my case it looked something like this:

image

As you can see it is taking 420ms to open a single connection to the Analysis Server This should be around 20ms or less. One of the annoying this about reporting services is it will open a fresh connection for every dataset, so with more complex reports this became unusable very quickly with some seemingly simple reports taking 30 seconds plus.

We used and xPath query and another report to visualise this performance data per data

image

which facilitates a report showing time spent between connection and actual query time

image

 

What was the cause

After trying it on various servers, laptops, desktops we managed to isolate it down to any of our SSAS Servers running inside Hyper-V on Windows 2012R2. Some further triage isolated the issue down to VMs which were using a broadcom network card with a known issue with Hyper-V and the NIC

http://www.dell.com/support/article/us/en/19/SLN132131/EN

https://support.microsoft.com/en-ie/kb/2902166

After some reconfiguration, the connections came down to under 20ms Smile

Analysing Report Server Logs

by Bob Duffy 4. March 2015 08:54

Attached is the example Power Pivot Model I used to analyse Report Server Logs complete with 3D pie chart to shock the DataViz world ;-)

I want to get time to extend this to include a lot of the “version 3” data which includes per Dataset metrics on query and processing time.

Feel free to use this as a starting point for analysing Report Server performance or usage.

image

You’ll need to update the connection in the Power Picot window to point to your ReportServer database.

 

ReportServerLogs.xlsxReportServerLogs.xlsx

ReportServerLogs.xlsxReportServerLogs.xlsx

Speaking Schedule First Half 2015

by Bob Duffy 1. February 2015 20:01

Quite a selection of speaking events coming up and an ambitious selection of mostly new content this year.

Belfast SQL User Group - Thurs 26th Feb.

  • When Good SQL Design Goes Bad
  • Optimising Reporting Services Performance

sqlBits, Excel London Exhibition Centre – 4th to 7th March

Friday 11:00 – 12:00 When Good SQL Design Goes Bad

Saturday 11:15 – 13:15 Scaling the DVD Store from 100 to a million TPM

Saturday  16:00 – 17:00 The European Economic Crisis And the Euro – A Data Tale

SQL Saturday Exeter. 24th –25th April, 2015

Friday Pre con Optimising Reporting Services for MDX Data Sources

Saturday 16:20 When Good SQL Design Goes Bad

Note sure if I will be selected, but I have submitted for SQL Saturday Edinburgh (BI Edition)  on June 13th. Edinburgh’s a city close to my heart with my college years spent there and Sandra living out there now.

And of course don’t forget SQL Saturday Dublin on the 20th/21st June. I may make an appearance on the lightning talks maybe but not a regular session as soo much to organise and so many great speakers submitted.

BTW Myself and Marco have some mad plans for a half day or whole day FREE session in Ireland too with myself and some other speakers . Will post when we get dates firmed up. Watch the UG newsletters and blogs

If you are looking at running SQL Server in the cloud a key question you will ask is “Should I rent or Buy” SQL licenses. E.g. should you use the stock images and pay by the hour or should you buy your own licences.

Some facts that will help you decide:

  • SQL Server Enterprise Edition can vary in price from say 16,000 euro to 24,000 euro per four cores depending on your licensing scheme and bartering ability.
  • You MUST pay for software assurance at 20% per year to transfer licenses to the cloud (Azure or Amazon)
  • Azure charges about 1,164 euro per month for enterprise edition, per four cores.
  • Buying licences with SA gives you rights to use a “DR” instance without paying for its license. When renting you need to pay for DR while its turned on (not good for say a passive mirror).

So assuming full retail prices we can say the following:

  • If you are running for more than 2 years at 100% up time you should buy.
  • If you are powering down images then factor this in. So if 50% powered down,  only buy if you are sweating the licences for > 4 years.
  • If you are running “warm” DR Servers then buy licenses if run for more than a year. Effectively buying licences becomes twice as appealing.

One thing I am looking at the moment is a scale out architecture where there may be two nodes up 100% of the time and then additional nodes scaled on demand. In this hybrid case it makes sense to buy licenses on the permanent nodes and “rent” on the elastic nodes.

Obviously this doesn’t take into account that renting gives you more flexibility, so is certainly the mode of choice for temporary environments.

Hot of the press is the announcement of MS Azure support for local SSD, up to 800GB which should be enough for “most” TempDBs.
http://azure.microsoft.com/blog/2014/09/22/new-d-series-virtual-machine-sizes/

Now this doesn’t have me too excited as its volatile and core storage is still a bit challenging to get IOPS on, especially if you need “burst” capability. What has me super excited is the introduction of the “memory” intensive images, or the D series with very competitive pricing.

clip_image002_thumb2

Annoyingly the pricing is a bit misleading as provisioning page has the price above (768.48 per month), but the pricing page has a different price

http://azure.microsoft.com/en-us/pricing/details/virtual-machines/

image_thumb[1]

 

If we look at a 16 core server options with close to 128GB ram we can compare the old and new offerings against Amazon. The only way to get 16 cores before was via the A9 image and it was really pricey, now we have local SSD thrown in for 51% price drop (ok with no infiniband, but I'm good with the trade).

  Azure (Old Pricing) Azure (new) Amazon EC2
Name A9 STANDARD_D14 r3.4xlarge
Cost/Month € 2,715 1,314 1,092
Spec 16 Cores
112 GB
40 GBit Infiniband
16 Cores
112 GB
800GB Local SSD
16 Cores
122 GB
2x320GB Local SSD

So, while the price for 16 core images has come down a lot its still not as competitive as the Amazon offering.

What is it missing to be a technically superior solution to Amazon for running SQL Server. A few things IMO:

  • Provisioned IOPS
  • Wholesale Replacement of old fashioned Magnetic Media with SSD
  • Burst Capably on IOPS

Amazon still has these trump cards and while it does, it will be holding the technical edge for running SQL Server at the high end on demanding IO based workloads.

However todays announcement is a huge leap forward. If you can leverage the Local SSD and/or don’t need massive IOPS then you can pocket a higher business value.

Why Pay for Professionals ?

by Bob Duffy 9. September 2014 09:23

Often, especially when working for larger organisations that need to go through a procurement process I get asked why our rates seem so high. They can hire SQL people from the global outsourced companies at 3-4 times less, and even go to general local consultancy practises for 1/2 of the rates.

The preferred solution to a lot of this trouble is to move away from charging by hour and to charge for results. Its super hard to do a lot of the time when people are so used to watching per hour costs rather than the quality and overall business value.

Here are two of my favourite memes doing the rounds.

“There is Always Someone who will do it Cheaper”

 

Another favourite line floating around the last while is

“If you think hiring a professional is expensive, wait until you hire an amateur”

Load Testing SSAS in Oslo

by Bob Duffy 30. August 2014 09:47

Thanks for everyone who made it to my SQL Saturday Session in Olso today on load testing Analysis Services.

If you are interested in getting started to building a load test harness in visual studio, I have attached the source code I used for the demos and sample database for AdventureWorks.

image

Happy Load Testing!

When you need to do load testing for Analysis Services its common to need to capture and “replay” MDX queries. I’ll go into replay tools later – the first step is to capture the queries. You will need

  • An XMLA trace definition file capturing at a minimum the QueryBegin and QueryEnd events (Events 9 and 10)
  • AN XMLA command to end the trace
  • The ASCMD.EXE utility to automate the starting/stopping of the trace. The utility is available here
    http://msftdbprodsamples.codeplex.com/downloads/get/384553

Here’s a quick video demonstrating the process of capturing a trace and importing the results into a data table.

More details on managing server side traces are also provided on this link here

http://blogs.msdn.com/b/pfebi/archive/2013/02/12/managing-server-side-analysis-services-profiler-trace.aspx

Page List

Page List