Category: SIEM

A SIEM Odyssey: How Albert Einstein Would Have Designed Your SIEM Architecture

Albert Einstein taught us that there are four dimensions: the three physical dimensions plus time. The light being generated by the sun exists, but it will take about eight minutes to reach the earth before it exists in our environment. Many of the lights you see in the sky at night were generated by stars millions of years ago, and may no longer exist today.

The four dimensions of spacetime can teach us a lot about the universe, and now a good lesson on SIEM architecture design. In SIEM environments, log data is sent through various layers, introducing a delay between the data source and destination. In a properly designed SIEM architecture, the delay between the source and destination should be minimal, a few minutes at most. But in an undersized SIEM architecture, delays between the source and destination can be high, and in worst cases, data may not reach the destination at all.

SIEM environments have three main layers. The first is the data sources, the various Windows servers, firewalls, and security tools that will either send data to your SIEM, or your SIEM will pull data from. The second layer is the Processing Layer, which consists of applications (Connectors/Forwarders/Ingestion Nodes) designed to process and structure log data, and forward it. The final layer is the Analytics Layer, which is where log data is stored, security analytics is performed, and end users search for data.

To highlight the risk introduced into your organization by an undersized SIEM architecture, we’ll use a DDoS attack against your organization as an example. The Bad Guyz Group has found a clever way to funnel millions out of your organization. Before they initiate the fraud scheme, they want to distract your organization from what is actually going on in order to buy themselves time, and thus launch a large-scale DDoS against your web servers.

Your DDoS protections begin sending out alerts, notifying your SOC of the attack and that there’s no concern at the moment. The amount of traffic directed at your web servers seems to be increasing, but is not near a level that will void your DDoS protections. Your SOC notifies leadership that even though there’s an active attack in progress, there’s nothing to be worried about.

While your DDoS protection is working as expected, your SIEM Processing Layer is being flooded with a 400% increase in firewall and proxy traffic. Your SIEM Processing Layer was only designed to process 10,000 events per second (EPS), and is now struggling to process a surge of 40,000 EPS. Cache files start to appear within minutes, growing at a rate of 100GB per hour, which will exhaust cache space within eight hours.

Hours later, SOC Analysts notice that the timestamps on most of the log data are several hours behind. They send an email to the SIEM Engineers, asking to see if there’s anything wrong with the SIEM application. When the SIEM Engineers get out of their project meeting a few hours later, they login to the servers and notice the cache files, extremely high EPS rates, and maxed-out RAM/CPU usage. They then notice that the surge in data is being caused by firewall and proxy logs. After a conversation with the SOC, the SIEM Engineers are then informed of the DDoS attack that happened earlier in the day.

Later in the evening, the SIEM Engineers warn leadership that the Processing Layer is dropping cache files of log data and is refusing new connections, resulting in data loss. The average log data delay now stands at eight hours as the DDoS attack continues.

Fortunately, the DDoS attack stopped the following morning, and the SIEM Processing Layer began reducing the amount of cache files on the servers. The SIEM Engineers anticipate that cache files should be completely cleared by 5PM.

Later that morning, the SOC Manager gets a call from the Fraud team, asking if they can see traffic to several IP Address. The SOC Analysts begin searching, but the response times are very slow and the latest data available is from last night. Just as the SIEM Engineers expected, by 5PM all cache files were cleared, and analysts were searching data in real-time again. They found only one hit from one of the IPs provided. The Fraud team insists there should be more than that, but the SIEM Engineers note that the other hits may have been dropped when the Processing Layer was refusing new connections during the surge in data. Leadership isn’t happy, and calls for an immediate review of the SIEM environment.

The bad news is that many SIEM environments are not sized appropriately to deal with such scenarios, or with legitimate data surges in general. These situations can leave your organization blind to an attack in progress, as the data required for an investigation exists, but is not yet available to your analysts, or worse is being dropped from existence.

The good news is that you can significantly reduce the probability and severity of this scenario. While SIEM environments can be expensive, the costly part is typically the Analytics Layer, and for many organizations over-sizing this layer isn’t an option. However, the Processing Layer tends to be much less expensive, and in some cases would simply result in the cost of the physical or virtual servers required.

A SIEM Processing Layer should be significantly larger than your sustained average event per second rate. While this number is a requirement to determine SIEM application licensing costs, many organizations make the mistake of sizing their SIEM according to this metric. In addition to spikes, the amount of traffic received by your SIEM during the day is likely to be much higher than at night. If your sustained EPS rate is 20,000 EPS, then it can be possible for your EPS during the day to be 30,000, and EPS at night to be in the 5,000-10,000 EPS range. If you receive a spike in traffic during the day, the 30,000 EPS can turn into 60,000 EPS. In many SIEM environments, this would cause the Processing Layer to quickly exhaust caches and begin dropping data. Supporting a large spike in traffic could simply be done by adding more devices (Connectors/Forwarders/Ingestion Nodes) within the Processing Layer. The increase in processing power and overall cache availability would reduce the risk of log transmission delays and data dropping.

In addition to reducing the probability of data delay and loss, an over-sized Processing Layer brings high availability benefits, and as well can make migrations and upgrades easier. If you have a single point of failure, you can lose your Processing Layer entirely if the device fails. If you have enough Connectors/Forwarders only to process your sustained EPS rate, you risk the above scenario when one of the devices within the Processing Layer fails, as the others have to make up for the extra EPS rates. If you need to upgrade your Processing Layer, the extra devices can make the upgrade smoother and transparent to any operational issues.

While we may have solved the issue with the Processing Layer, the surge in data can also result in transmission delays, data loss, slower end-user search response times, and system stability issues on the Analytics Layer. However, while it may seem logical to build an Analytics Layer to support double the sustained EPS rates, it can be cost prohibitive for many organizations. An adequately-sized Processing Layer can assist during surges by aggregating data (combining similar events into one, for SIEM products that support aggregation), caching it locally, and limiting the EPS-out rates to the Analytics Layer. Your SIEM Engineers can also control what data is sent over others, so if there’s a dire need for a particular data source, your staff can limit other data sources to be sent to the Analytics Layer so that the pertinent sources can be sent in priority.

In summary, there’s a strong return on investment for building an adequate SIEM Processing Layer given the low costs, risk reduction, and invaluable security benefits. Even with a minimal Analytics Layer, a properly-sized Processing Layer will be sufficiently able to process a surge of data, cache data that can’t be forwarded, prevent data loss, and reduce risks caused by large increases in log data. Make the investment and leave spacetime issues for the physicists!

Please like, share, or comment if you enjoyed this article. Thank you!

The Pros and Cons of Structuring Log Data at Ingestion Time with SIEMs

Another important but often overlooked part of a SIEM architecture and design or product analysis exercise is whether the product structures the data or not as it’s ingested/processed, and how that can affect your organization’s environment. This seemingly miniscule functionality can have a significant effect on your SIEM environment, and can even introduce risk into your organization.

Let’s start with the advantages of structuring (parsing) log events at ingestion time. In general, structuring data as its ingested/processed can give you many opportunities to manipulate the data in a positive way.

Advantages of Structuring Log Data at Ingestion Time

1. Ability to Aggregate

Aggregation gives the SIEM tool the ability to combine multiple, similar events into one single event. The biggest advantage of this is the reduction in EPS rates processed by the SIEM and the reduced storage requirements. Ten firewall deny events from the same source and destination can be combined into one, using 1,000 bytes of storage instead of 10,000 bytes. SIEM tools typically have the ability to aggregate hundreds of events over a period of thirty seconds or more, so it’s common to see aggregation successfully combine hundreds of events into one. Caution must be used with aggregation settings as the higher the aggregation window (the maximum amount of time the Connector/Log Processor will wait for similar events to combine into one), the higher the memory requirement for it.

2. Ability to Standardize Casing

Most SIEM tools can easily standardize casing of all fields parsed by the Connector/Log Processor. This is often an overlooked benefit that comes with structuring data, as the various data sources in your environment will log in various casings, and thus introduce a potential security risk. The security risk that can be introduced is that your security analysts may be getting null search results when the data they need is actually there.

In several investigations, SOC analysts were having issues getting hits for a particular user’s data. Upon closer inspection, we found the desired data, and discovered that the initial searches were coming up null because the casing in the searches did not match that of the log data. The SOC analysts were searching for “frank,” but the SIEM tool was configured to be case-sensitive, and the Windows logs were being logged in uppercase as “FRANK.” Thus, by simply having the Connector standardize casing for particular fields, you can minimize the above scenarios.

Standardizing casing can also increase search performance. When the SIEM tool only has to search in one case, it reduces the amount of characters it needs to search for. A simple search for “FRANK” only requires the tool to match on five characters instead of thirty-two (FRANK, frank, Frank, FRank, FRAnk, etc).

Why not simply disable case-sensitivity for searches? This is seemingly the best option, as the risk of missing data is mitigated. However, the major disadvantage of this option is that it increases the processing power required for the searches. For smaller environments, this is practical and the effects will likely be negligible. However, in larger environments where the SIEM is processing several thousand events per second and there are several end users, the results can be noticeable.

A practical work around can be to disable case-sensitivity for particular searches at the discretion of the analyst. Many SIEM tools offer this option for this very reason. There’s typically an option before the search to disable case-sensitivity.

A best practice to mitigate missing data due to casing issues while maximizing performance is to start with a generic case insensitive search, and once you get hits on the data you’re searching for, switch to the casing you see. For example, if you’re looking for user Frank’s Windows logs, start with a small, e.g. few minute, case-insensitive search “frank”, and once you see that the Windows logs are logging it as “Frank,” switch to the proper casing and then expand the search. This is a practical option that will help analysts avoid missing data, and will not require you to configure your tool to be case insensitive.

Regardless of how you chose to configure case-sensitivity, simply ensure your staff understand how your environment works and best practices for searching your data.

3. Ability to Add and Modify Fields

Many SIEMs can append data to existing fields, override fields with new data, and modify values put into fields. A common nuisance when searching for log data is how some systems have their FQDN logged (e.g. while others simply log the server name (server01). This can cause a similar risk as case-sensitivity, where SOC analysts search for Device Host Name =”server01” but get no results as the server appears in the logs as “”. This forces the SOC analyst to do a wildcard search of Device Host Name =”server01*” etc, and ultimately requires more processing power from the SIEM.

When data is structured/parsed at ingestion time, the SIEM can do a simple lookup of the e.g. first period, strip whatever proceeds the period, and then put that into another field. Using “” as an example, the parser can leave server01 in the host name field, and then put the stripped in the e.g. domain field. Analysts then know that they only need to search for the server name in the host name field, and to search the domain field if they want to know the domain of the server.


Now that we’ve fallen in love with the advantages of structuring data at ingestion time, let’s look at the disadvantages before we leave for the honeymoon.

The Disadvantages of Structuring Log Data at Ingestion Time

1. Increased Event Size

The first, and potentially most costly aspect of structuring the data at ingestion time is the increase in event size. When you structure the data, you increase the size of the event, in many cases doubling its size or more. Please see my related article, The Million Dollar SIEM Question: To Parse or Not To Parse, for more on this.

2. Potential Data Loss and Integrity Issues

Given that your parser is instructed to place values in specific fields, for example taking the value after the second comma and putting it into the user name field, there’s potential integrity and data loss issues if your parser is not updated at the same time the logging format for a particular data source changes.

Let’s take a look at a sample log event:

01-11-2018 14:12:22,, frank, authentication, interactive login, successful
The parser takes the timestamp from the characters before the first comma, the IP Address after the first comma, the user name after the second comma, the type of event after the third comma, the type of login after the fourth comma, and finally the outcome after the fifth comma. All is well until the vendor decides to add a new field, an event code, and change the order of the events:

01-11-2018 14:12:22, 4390000,, authentication, interactive login, successful, frank

This is a simple change, but your parser needs to be updated to ensure that the values are put into the correct fields, and to add in the new field. Should this new log type be implemented without a corresponding parser change, we’re not only going to have data in the wrong fields, we’re not going to know who did the login, as the value “frank” will not be visible to the parser.

3. Increased System Requirements

The more modifications the parser has to make to the event, the more processing power the Connector/Forwarder/Processor will consume. Ensure there are sufficient system resources able to process the required modifications.


A Summary of the Pros and Cons of Structuring data at ingestion time:

The Million Dollar SIEM Question: To Parse or Not To Parse

Given that SIEMs process and store data, one of the major requirements of a successful SIEM environment is proper and sufficient storage. Depending on your organization’s SIEM requirements, the cost of storage alone for your SIEM can exceed application licensing costs.

The most common omission in a SIEM product selection analysis is differentiating how the proposed applications process and store data. SIEMs process and store data differently, and thus will all produce different storage requirements. Given that storage is a major cost of your environment, how the application processes and stores data can alter your storage costs significantly.

Traditional, as well as newer SIEM products are designed to parse data as it’s ingested, and thus store data as a parsed event. Additionally, SIEM products will parse the data into different field sets; a Windows event in SIEM Product A can be parsed into 200 fields, while SIEM Product B will parse it into 193 fields. This will result in a different event size for the same data. Some newer SIEM products do not parse data as it’s ingested, stores the data raw, and only parses it when required, e.g. when you run a query, report, etc.

There are advantages and disadvantages to both. Parsing (or normalizing/enhancing/enriching) the data structures it, by adding applicable metadata. For example, the following log entry ‘jsmith,, failed login’ would appear parsed as ‘username=jsmith, IPAddress=,event name=failed login’. While this makes the data more organized and allows for more refined searches, it makes the log entry bigger in size. A 500 byte raw event can turn into a 1000 byte parsed event, doubling the storage requirement for this event.

While parsing increases the size of the event, the SIEM tool does gain the ability to manipulate the data. Many SIEMs that structure data have the ability to aggregate events, which can take multiple, similar events and combine them into one. For example, the event ‘source address=, source port=9022, event name= deny, destination address =, destination port= 443’ that occurs 10 times can be combined into one, with an extra field added, e.g. event count =10, to indicate how many times the event occurred. Thus, 10 events at 1000 bytes each use 1000 bytes of storage with the structured SIEM application instead of 10,000 bytes.

To highlight how this can affect your organization, let’s look at Company A as an example. The technical staff at Company A have been reading about the benefits of SIEM from some guy on the Internet, and decide they want to implement one. They want 2 months of online data followed by 10 months offline, and for the data to be highly available. Their storage team can provide high-speed storage for $5,000 per terabyte, and lower-speed storage for $2,000 per terabyte.

Next, they’ve invited a couple of vendors in for a product overview, and they’re making each complete the SIEM Storage Requirements and Costs spreadsheet they created.

First up for Company A is Product A. Product A does not parse event data as it’s ingested and stores logs in raw format. It gets up to 50% compression on live data, and 85% compression on archived data. The product can replicate data and thus easily meet the high availability requirement. The tech staff at Company A also went the extra mile and determined that the average log event size from all their systems is 700 bytes.

A summary of the requirements and weights so far:

Again, since Product A stores data only in raw format, the Average Normalized Event Size is 0, as it does not store a parsed/normalized event. Product A cannot aggregate data, so there is no Aggregation Benefit. The product will create a copy of each event, bringing the replication factor to 2.

Next, we’re going to determine the average sustained events per second rate from the numbers the tech staff provided. Based on the total number of devices, the SIEM will need sufficient storage to store 5,000 Total Average EPS (events per second) at 700 bytes per event. Using the SIEM Storage Requirements and Costs spreadsheet, we get the following table.

The total daily uncompressed storage requirement is 607 GB. At 50% compression, the Total Online Storage Requirement is 18 TB. At 85% compression, the Total Offline Storage Requirement is 28TB. 18 TB at $5,000 per TB brings the cost to $91,000, and 28 TB at $2,000 adds another $55K, bringing the total to approx. $146,000.

Up next is Product B, which is a SIEM tool that’s designed to work with structured data. It will parse log events and create an Average Normalized Event Size of 2,000 bytes, and will be able to reduce events by 40% through aggregation. It can’t replicate data, but the Connector/Forwarder will be configured to send to 2 destinations. And by complete chance, it has the exact same compression ratios.

Product B is going to process the exact same raw EPS, but the Aggregation Benefit drops the sustained Total Average EPS to 3,000.

The total daily uncompressed storage requirement is approx. 1 TB. At 50% compression, the total Online Storage Requirement is 31.2 TB. At 85% compression, the total Offline Storage Requirement is 47.6 TB. 31.2 TB at $5,000 per TB brings the cost to $156,000, and 47.6 TB at $2,000 adds another $95K, bringing the total to approx. $251,000.

As you can see, for the sample insurance company, Product A produces a significantly different storage cost than Product B due to the way the products process and store data. The tech staff at Company A like Product B better, but know that it will be tough to sell the VPs a solution that will cost $500,000 more over the next five years.

However, the result could be completely different at Company B, which has different requirements than Company A. Company B is a telco that wants their IT staff to monitor their network infrastructure with a SIEM. The tech staff at Company A were generous enough to share the SIEM Storage Requirements and Costs spreadsheet with their buddies at Company B.

Company B decides to bring in Product A first and has them fill out the spreadsheet. As the environment mainly consists of network devices, the Average Raw Event Size is going to be 450 bytes. Again, since Product A doesn’t parse log data, there is no Average Normalized Event Size or Aggregation Benefit.

Next, we can calculate a total daily uncompressed storage requirement of 1.2 TB from the 15,780 EPS rate. That will produce an online storage requirement of 37 TB and offline of 56 TB.

That will bring the total yearly storage cost to approx. $300,000.

Next, the tech staff at Company B bring in Product B for an overview.

The mostly network devices will produce an Average Normalized Event Size of 1500 bytes, and the Aggregation Benefit will be very high for the firewall data, reaching an overall total benefit estimated at 80%.

Through the strong aggregation benefit, the total EPS is reduced from nearly 16,000 to just over 3,000. As a result, the storage requirements for Product B are 24.5 TB for online, and 37.4 TB for offline, respectively.

So for Product B, the yearly storage costs sit at $200,000 per year.

The tech staff at Company B like Product A better for some reason, but like at Company A, they know it will be tough to justify the extra $500,000 in storage costs over the next five years.

So as you can see, To Parse Or Not To Parse is not some Shakespearean-sounding cliché by some geek on the Internet. Requirements such as storing two copies of each event, storing both the raw and normalized event (add the storage costs for Product A and B together, roughly!), or retaining two years of log data can create tremendous storage cost differences over the tenure of your SIEM environment.

The best product for your company will be that which meets your requirements best, and storage costs alone should not be the deciding factor for which product is selected for your organization. There are many advantages and disadvantages to leaving data in raw format or parsing it, but as the data shows, you can’t ignore a question that can really be in the millions!


For the record, the lads at Company A have shared the SIEM Storage Requirements and Costs spreadsheet.