共用方式為

A few thoughts on sizing Audit Collection System

  • 發行項

People were already collecting logs with MOM, so why not the security log? Some people were doing that, but it did not scale enough; for this reason, a few years ago Eric Fitzgerald announced that he was working on Microsoft Audit Collection System. Anyhow, the tool as it was had no interface… and the rest is history: it has been integrated into System Center Operations Manager. Anyhow, ACS remains a lesser-known component of OpsMgr.

There are a number of resources on the web that is worth mentioning and linking to:

and, of course, many more, I cannot link them all.

As for myself, I have been playing with ACS since those early beta days (before I joined Microsoft and before going back to MOM, when I was working in Security), but I never really blogged about this piece.

Since I have been doing quite a lot of work around ACS lately, again, I thought it might be worth consolidating some thoughts about it, hence this post.

Anatomy of an “Online” Sizing Calculation

What I would like to explain here is the strategy and process I go thru when analyzing the data stored in a ACS database, in order to determine a filtering strategy: what to keep and what not to keep, by applying a filter on the ACS Collector.

So, the first thing I usually start with is using one of the many “ACS sizer” Excel spreadsheets around… which usually tell you that you need more space than it really is necessary… basically giving you a “worst case” scenario. I don’t know how some people can actually do this from a purely theoretical point of view, but I usually prefer a bottom up approach: I look at the actual data that the ACS is collecting without filters, and start from there for a better/more accurate sizing.

In the case of a new install this is easy – you just turn ACS on, set the retention to a few days (one or two weeks maximum), give the DB plenty of space to make sure it will make it, add all your forwarders… sit back and wait.

Then you come back 2 weeks later and start looking at the data that has been collected.

What/How much data are we collecting?

First of all, if we have not changed the default settings, the grooming and partitioning algorithm will create new partitioned tables every day. So my first step is to see how big each “partition” is.

But… what is a partition, anyway? A partition is a set of 4 tables joint together:

  1. dtEvent_GUID
  2. dtEventData_GUID
  3. dtPrincipal_GUID
  4. dtSTrings_GUID

where GUID is a new GUID every day, and of course the 4 tables that make up a daily partition will have the same GUID.

The dtPartition table contains a list of all partitions and their GUIDs, together with their start and closing time.

Just to get a rough estimate we can ignore the space used by the last three tables – which are usually very small – and only use the dtEvent_GUID table to get the number of events for that day, and use the stored procedure “sp_spaceused”  against that same table to get an overall idea of how much space that day is taking in the database.

By following this process, I come up with something like the following:

Partition ID Status Partition Start Time Partition Close Time Rows Reserved KB Total GB
9b45a567_c848_4a32_9c35_39b402ea0ee2 0 2/1/2010 2:00 2/1/2010 2:00 29,749,366 7,663,488 7,484
8d8c8ee1_4c5c_4dea_b6df_82233c52e346 2 1/31/2010 2:00 2/1/2010 2:00 28,067,438 9,076,904 8,864
34ce995b_689b_46ae_b9d3_c644cfb66e01 2 1/30/2010 2:00 1/31/2010 2:00 30,485,110 9,857,896 9,627
bb7ea5d3_f751_473a_a835_1d1d42683039 2 1/29/2010 2:00 1/30/2010 2:00 48,464,952 15,670,792 15,304
ee262692_beae_4d81_8079_470a54567946 2 1/28/2010 2:00 1/29/2010 2:00 48,980,178 15,836,416 15,465
7984b5b8_ddea_4e9c_9e51_0ee7a413b4c9 2 1/27/2010 2:00 1/28/2010 2:00 51,295,777 16,585,408 16,197
d93b9f0e_2ec3_4f61_b5e0_b600bbe173d2 2 1/26/2010 2:00 1/27/2010 2:00 53,385,239 17,262,232 16,858
8ce1b69a_7839_4a05_8785_29fd6bfeda5f 2 1/25/2010 2:00 1/26/2010 2:00 55,997,546 18,105,840 17,681
19aeb336_252d_4099_9a55_81895bfe5860 2 1/24/2010 2:00 1/24/2010 2:00 28,525,304 7,345,120 7,173
1cf70e01_3465_44dc_9d5c_4f3700dc408a 2 1/23/2010 2:00 1/23/2010 2:00 26,046,092 6,673,472 6,517
f5ec207f_158c_47a8_b15f_8aab177a6305 2 1/22/2010 2:00 1/22/2010 2:00 47,818,322 12,302,208 12,014
b48dabe6_a483_4c60_bb4d_93b7d3549b3e 2 1/21/2010 2:00 1/21/2010 2:00 55,060,150 14,155,392 13,824
efe66c10_0cf2_4327_adbf_bebb97551c93 2 1/20/2010 2:00 1/20/2010 2:00 58,322,217 15,029,216 14,677
0231463e_8d50_4a42_a834_baf55e6b4dcd 2 1/19/2010 2:00 1/19/2010 2:00 61,257,393 15,741,248 15,372
510acc08_dc59_482e_a353_bfae1f85e648 2 1/18/2010 2:00 1/18/2010 2:00 64,579,122 16,612,512 16,223

If you have just installed ACS and let it run without filters with your agents for a couple of weeks, you should get some numbers like those above for your “couple of weeks” of analysis. If you graph your numbers in Excel (both size and number of rows/events per day) you should get some similar lines that show a pattern or trend:

Trend: Space user by day

Trend: Number of events by day

So, in my example above, we can clearly observe a “weekly” pattern (monday-to-friday being busier than the weekend) and we can see that – for that environment – the biggest partition is roughly 17GB. If we round this up to 20GB – and also considering the weekends are much quieter – we can forecast 20*7 = 140GB per week. This has an excess “buffer” which will let the system survive event storms, should they happen. We also always recommend having some free space to allow for re-indexing operations.

In fact, especially when collecting everything without filters, the daily size is a lot less predictable: imagine worms “trying out” administrator account’s passwords, and so on… those things can easily create event storms.

Anyway, in the example above, the customer would have liked to keep 6 MONTHS (180days) of data online, which would become 20*180 = 3600GB = THREE TERABYTE and a HALF! Therefore we need a filtering strategy – and badly – to reduce this size.

[edited on May 7th 2010 - if you want to automate the above analysis and produce a table and graphs like those just shown, you should look at my following post.]

Filtering Strategies

Ok, then we need to look at WHAT actually comprises that amount of events we are collecting without filters. As I wrote above, I usually run queries to get this type of information.

I will not get into HOW TO write a filter here – a collector’s filter is a WMI notification query and it is already described pretty well elsewhere how to configure it.

Here, instead, I want to walk thru the process and the queries I use to understand where the noise comes from and what could be filtered – and get an estimate of how much space we could be saving if filter one way or another.

Number of Events per User

--event count by User (with Percentages)
declare @total float
select @total = count(HeaderUser) from AdtServer.dvHeader
select count(HeaderUser),HeaderUser, cast(convert(float,(count(HeaderUser)) / (convert(float,@total)) * 100) as decimal(10,2))
from AdtServer.dvHeader
group by HeaderUser
order by count(HeaderUser) desc

In our example above, over the 14 days we were observing, we obtained percentages like the following ones:

#evt HeaderUser Account Percent
204,904,332 SYSTEM 40.79 %
18,811,139 LOCAL SERVICE 3.74 %
14,883,946 ANONYMOUS LOGON 2.96 %
10,536,317 appintrauser 2.09 %
5,590,434 mossfarmusr …

Just by looking at this, it is pretty clear that filtering out events tracked by the accounts “SYSTEM”, “LOCAL SERVICE” and “ANONYMOUS”, we would save over 45% of the disk space!

Number of Events by EventID

Similarly, we can look at how different Event IDs have different weights on the total amount of events tracked in the database:

--event count by ID (with Percentages)
declare @total float
select @total = count(EventId) from AdtServer.dvHeader
select count(EventId),EventId, cast(convert(float,(count(EventId)) / (convert(float,@total)) * 100) as decimal(10,2))
from AdtServer.dvHeader
group by EventId
order by count(EventId) desc

We would get some similar information here:

Event ID Meaning Sum of events Percent
538 A user logged off 99,494,648 27.63
540 Successful Network Logon 97,819,640 27.16
672 Authentication Ticket Request 52,281,129 14.52
680 Account Used for Logon by (Windows 2000) 35,141,235 9.76
576 Specified privileges were added to a user's access token. 26,154,761 7.26
8086 Custom Application ID 18,789,599 5.21
673 Service Ticket Request 10,641,090 2.95
675 Pre-Authentication Failed 7,890,823 2.19
552 Logon attempt using explicit credentials 4,143,741 1.15
539 Logon Failure - Account locked out 2,383,809 0.66
528 Successful Logon 1,764,697 0.49

Also, do not forget that ACS provides some report to do this type of analysis out of the box, even if for my experience they are generally slower – on large datasets – than the queries provided here. Also, a number of reports have been buggy over time, so I just prefer to run queries and be on the safe side.

Below an example of such report (even if run against a different environment – just in case you were wondering why the numbers were not the same ones :-)):Event Counts ACS Default Report

The numbers and percentages we got from the two queries above should already point us in the right direction about what we might want to adjust in either our auditing policy directly on Windows and/or decide if there is something we want to filter out at the collector level (here you should ask yourself the question: “if they aren’t worth collecting are they worth generating?” – but I digress).

Also, a permutation of the above two queries should let you see which user is generating the most “noise” in regards to some events and not other ones… for example:

--event distribution for a specific user (change the @user) - with percentages for the user and compared with the total #events in the DB
declare @user varchar(255)
set @user = 'SYSTEM'
declare @total float
select @total = count(Id) from AdtServer.dvHeader
declare @totalforuser float
select @totalforuser = count(Id) from AdtServer.dvHeader where HeaderUser = @user
select count(Id), EventID, cast(convert(float,(count(Id)) / convert(float,@totalforuser) * 100) as decimal(10,2)) as PercentageForUser, cast(convert(float,(count(Id)) / (convert(float,@total)) * 100) as decimal(10,2)) as PercentageTotal
from AdtServer.dvHeader
where HeaderUser = @user
group by EventID
order by count(Id) desc

The above is particularly important, as we might want to filter out a number of events for the SYSTEM account (i.e. logons that occur when starting and stopping services) but we might want to keep other events that are tracked by the SYSTEM account too, such as an administrator having wiped the Security Log clean – which might be something you want to keep:

Event ID 517 Audit Log was cleared

of course the amount of EventIDs 517 over the total of events tracked by the SYSTEM account will not be as many, and we can still filter the other ones out.

Number of Events by EventID and by User

We could also combine the two approaches above – by EventID and by User:

select count(Id),HeaderUser, EventId

from AdtServer.dvHeader

group by HeaderUser, EventId

order by count(Id) desc

This will produce a table like the following one

SQL Query: Events by EventID and by User

which can be easily copied/pasted into Excel in order to produce a pivot Table:

Pivot Table

Cluster EventLog Replication

One more aspect that is less widely known, but I think is worth showing, is the way that clusters behave when in ACS. I don’t mean all clusters… but if you keep the “eventlog replication” feature of clusters enabled (you should disable it also from a monitoring perspective, but I digress), each cluster node’s security eventlog will have events not just for itself, but for all other nodes as well.

Albeit I have not found a reliable way to filter out – other than disabling eventlog replication altogether.

Anyway, just to get an idea of how much this type of “duplicate” events weights on the total, I use the following query, that tells you how many events for each machine are tracked by another machine:

--to spot machines that are cluster nodes with eventlog repliation and write duplicate events (slow)

select Count(Id) as Total,replace(right(AgentMachine, (len(AgentMachine) - patindex('%\%',AgentMachine))),'$','') as ForwarderMachine, EventMachine

from AdtServer.dvHeader

--where ForwarderMachine <> EventMachine

group by EventMachine,replace(right(AgentMachine, (len(AgentMachine) - patindex('%\%',AgentMachine))),'$','')

order by ForwarderMachine,EventMachine

Cluster Events

Those presented above are just some of the approaches I usually look into at first. Of course there are a number more. Here I am including the same queries already shown in action, plus a few more that can be useful in this process.

I have even considered building a page with all these queries – a bit like those that Kevin is collecting for OpsMgr (we actually wrote some of them together when building the OpsMgr Health Check)… shall I move the below queries on such a page? I though I’d list them here and give some background on how I normally use them, to start off with.

Some more Useful Queries

--top event ids
select count(EventId), EventId
from AdtServer.dvHeader
group by EventId
order by count(EventId) desc

--event count by ID (with Percentages)
declare @total float
select @total = count(EventId) from AdtServer.dvHeader
select count(EventId),EventId, cast(convert(float,(count(EventId)) / (convert(float,@total)) * 100) as decimal(10,2))
from AdtServer.dvHeader
group by EventId
order by count(EventId) desc

--which machines have ever written event 538
select distinct EventMachine, count(EventId) as total
from AdtServer.dvHeader
where EventID = 538
group by EventMachine

--machines
select * from dtMachine

--machines (more readable)
select replace(right(Description, (len(Description) - patindex('%\%',Description))),'$','')
from dtMachine

--events by machine
select count(EventMachine), EventMachine
from AdtServer.dvHeader
group by EventMachine

--rows where EventMachine field not available (typically events written by ACS itself for chekpointing)
select *
from AdtServer.dvHeader
where EventMachine = 'n/a'

--event count by day
select convert(varchar(20), CreationTime, 102) as Date, count(EventMachine) as total
from AdtServer.dvHeader
group by convert(varchar(20), CreationTime, 102)
order by convert(varchar(20), CreationTime, 102)

--event count by day and by machine
select convert(varchar(20), CreationTime, 102) as Date, EventMachine, count(EventMachine) as total
from AdtServer.dvHeader
group by EventMachine, convert(varchar(20), CreationTime, 102)
order by convert(varchar(20), CreationTime, 102)

--event count by machine and by date (distinuishes between AgentMachine and EventMachine
select convert(varchar(10),CreationTime,102),Count(Id),EventMachine,AgentMachine
from AdtServer.dvHeader
group by convert(varchar(10),CreationTime,102),EventMachine,AgentMachine
order by convert(varchar(10),CreationTime,102) desc ,EventMachine

--event count by User
select count(Id),HeaderUser
from AdtServer.dvHeader
group by HeaderUser
order by count(Id) desc

--event count by User (with Percentages)
declare @total float
select @total = count(HeaderUser) from AdtServer.dvHeader
select count(HeaderUser),HeaderUser, cast(convert(float,(count(HeaderUser)) / (convert(float,@total)) * 100) as decimal(10,2))
from AdtServer.dvHeader
group by HeaderUser
order by count(HeaderUser) desc

--event distribution for a specific user (change the @user) - with percentages for the user and compared with the total #events in the DB
declare @user varchar(255)
set @user = 'SYSTEM'
declare @total float
select @total = count(Id) from AdtServer.dvHeader
declare @totalforuser float
select @totalforuser = count(Id) from AdtServer.dvHeader where HeaderUser = @user
select count(Id), EventID, cast(convert(float,(count(Id)) / convert(float,@totalforuser) * 100) as decimal(10,2)) as PercentageForUser, cast(convert(float,(count(Id)) / (convert(float,@total)) * 100) as decimal(10,2)) as PercentageTotal
from AdtServer.dvHeader
where HeaderUser = @user
group by EventID
order by count(Id) desc

--to spot machines that write duplicate events (such as cluster nodes with eventlog replication enabled)
select Count(Id),EventMachine,AgentMachine
from AdtServer.dvHeader
group by EventMachine,AgentMachine
order by EventMachine

--to spot machines that are cluster nodes with eventlog repliation and write duplicate events (better but slower)
select Count(Id) as Total,replace(right(AgentMachine, (len(AgentMachine) - patindex('%\%',AgentMachine))),'$','') as ForwarderMachine, EventMachine
from AdtServer.dvHeader
--where ForwarderMachine <> EventMachine
group by EventMachine,replace(right(AgentMachine, (len(AgentMachine) - patindex('%\%',AgentMachine))),'$','')
order by ForwarderMachine,EventMachine

--which user and from which machine is target of elevation (network service doing "runas" is a 552 event)
select count(Id),EventMachine, TargetUser
from AdtServer.dvHeader
where HeaderUser = 'NETWORK SERVICE'
and EventID = 552
group by EventMachine, TargetUser
order by count(Id) desc

--by hour, minute and user
--(change the timestamp)... this query is useful to search which users are active in a given time period...
--helpful to spot "peaks" of activities such as password brute force attacks, or other activities limited in time.
select datepart(hour,CreationTime) as Hours, datepart(minute,CreationTime) as Minutes, HeaderUser, count(Id) as total
from AdtServer.dvHeader
where CreationTime < '2010-02-22T16:00:00.000'
and CreationTime > '2010-02-22T15:00:00.000'
group by datepart(hour,CreationTime), datepart(minute,CreationTime),HeaderUser
order by datepart(hour,CreationTime), datepart(minute,CreationTime),HeaderUser