ZGrab performance issue with http?

[phillip-stephens]

Not a bug per se, but was investigating a slow zgrab scan and noticed that it maxes out at sending 6 MB/s. While there might be a lot of CPU overhead, this just seems egregiously slow and surely there's some performance gains we can achieve. This issue is to track that.

CLI command - cat ~/100k-domains.txt | ./zgrab2 http --max-redirects=3

This was run on a VM with a 1 Gb/s + link and plenty of cores(22)/RAM (56 GB)

[mzpqnxow]

Not a bug per se, but was investigating a slow zgrab scan and noticed that it maxes out at sending 6 MB/s. While there might be a lot of CPU overhead, this just seems egregiously slow and surely there's some performance gains we can achieve. This issue is to track that.

CLI command - cat ~/100k-domains.txt | ./zgrab2 http --max-redirects=3

This was run on a VM with a 1 Gb/s + link and plenty of cores(22)/RAM (56 GB)

Does using the "senders" global (-s if I remember correctly..) not help with this?

Regardless, it seems like a rather unusual number as a default... unless it's as a ratio to the number of logical cores. If that's the case, it does in fact seem to be very low, assuming your measurement is correct <-- misread as "6 connections", disregard this

In the past, users have created issues about performance that were ultimately caused by DNS lookups. I suspect that may be the case here. The spikes in the graph may represent time doing DNS lookups

I've managed to avoid DNS issues by pre-resolving targets, then using the targets file format with both the Virtual Host/FQDN and IP address, so DNS is only needed on redirects. This is necessary anyway if you want to cover all of the IP addresses for load-balanced FQDNs that return more than one IP in the A record lookup

In my opinion, zgrab2 should not be relied upon for DNS lookups, because of this exact issue - on the fly DNS lookups can significantly slow down the actual scanning/probes. I don't know if zmap project folks agree, but I've become pretty firm in believing it

[Seanstoppable]

I recently noticed that we do DNS lookups whenever we get a redirect, irrespective of if we are going to follow it, which is probably at least part of the problem.

[Seanstoppable]

I've tracked this to this line: https://github.com/zmap/zgrab2/pull/462/files#diff-dddc8e7eb09019004b9ea48a345b3fb794995c7b661e5e2aa0494d56d6ce7bd3L542 Where we go through and create a new request, parsing the DNS name and triggering a DNS lookup before evaluating whether we want to bail out due to redirects. Options I see are moving out the redirect check to right after the request lookup (per the PR), or I guess we could move it sooner in the loop, before we go to create a new request object/do the parsing.

[mzpqnxow]

I recently noticed that we do DNS lookups whenever we get a redirect, irrespective of if we are going to follow it, which is probably at least part of the problem.

Indeed, that's the residual bit that remains, even if you do provide all of the initial addresses. And I'm not sure of any way around it

Though I haven't actually measured it, I would assume it's a minor blip when compared with most target lists. Or maybe not, I haven't properly thought about it

[mzpqnxow]

I've tracked this to this line:

https://github.com/zmap/zgrab2/pull/462/files#diff-dddc8e7eb09019004b9ea48a345b3fb794995c7b661e5e2aa0494d56d6ce7bd3L542

Where we go through and create a new request, parsing the DNS name and triggering a DNS lookup before evaluating whether we want to bail out due to redirects.

Options I see are moving out the redirect check to right after the request lookup (per the PR), or I guess we could move it sooner in the loop, before we go to create a new request object/do the parsing.

One question- do you know off-hand what happens for a case like http://site.com redirect -> https://site.com?

Specifically, wondering if it knows that it can use that same address or it does another lookup

I imagine that is a good portion of the redirects for most target lists

[mzpqnxow]

One more note: if you do tackle this, keep in mind the potential pitfall mentioned in #307 which is really a general pattern that has manifested in different ways/places- TLS handshake failures, redirect exceeded, etc. I think your recent PR may have fixed this sort of behavior, just wanted to mention it

I'm not sure I ever sent a PR for the TLS handshake case 😞

[Seanstoppable]

My use case is a bit simpler, as a lot of my scans are just IPs and not domains So any redirect to a domain in that case does a DNS lookup In the case of cat ~/100k-domains.txt | ./zgrab2 http --max-redirects=3, you are going to do up to 5 lookups in the worst case: 1 for the initial domain, 3 for the up to 3 redirects, and a 4th if the last redirect also redirects, even if you don't follow it or use that data.

One question- do you know off-hand what happens for a case like http://site.com/ redirect -> https://site.com/?

I would expect it to just be cached locally, rather than zgrab2 trying to re-resolve everything. http://site.com/ -> http://www.site.com/ would trigger another lookup though.

However, I don't think the golang code will cache those internally, so maybe a small improvement if we cached those for the http -> https redirect we are expecting when we start at http, avoiding a roundtrip to the system resolver even if the value is cached there.

Indeed, that's the residual bit that remains, even if you do provide all of the initial addresses. And I'm not sure of any way around it

Yeah. I am not sure how to solve that one. Iterative process of doing one level at a time is a lot of wrapping/stitching together. I guess you could try to embed something like ZDNS inside (and maybe this is why there has been work to make it more library work) to do faster lookups than what we are doing today.

I wonder if part of it is also redirects like domain.com -> www.domain.com, where that doesn't take advantage of multiplexing. I wonder if you could be a little 'clever' and try to get the roundtripper to maintain a connection per IP address, rather than per a hostname.

[mzpqnxow]

My use case is a bit simpler, as a lot of my scans are just IPs and not domains So any redirect to a domain in that case does a DNS lookup In the case of cat ~/100k-domains.txt | ./zgrab2 http --max-redirects=3, you are going to do up to 5 lookups in the worst case: 1 for the initial domain, 3 for the up to 3 redirects, and a 4th if the last redirect also redirects, even if you don't follow it or use that data.

I see. So the initial request for our cases is similar (profile-wise) because no DNS lookup required, but your case won’t necessarily benefit at all from caching DNS from that first request (since there isn’t one)

Actually, let me correct myself. I just realized that my target list could actually prime a cache. Since I provide, for example:

1.2.3.4, target.com
3.4.5.6, www.target.com
…

… it should be possible to first reference the name to IP mappings in the target list, before making a DNS request. Wouldn’t help you, though. I wonder how much of an impact this would have for me. As a mapping with writes occurring only at startup, it wouldn’t need any locking. The lookups should be cheap. Just a matter of how many cases would actually get a hit, which would vary for each user’s target set

I would expect it to just be cached locally, rather than zgrab2 trying to re-resolve everything. http://site.com/ -> http://www.site.com/ would trigger another lookup though.

I would like to confirm that if I get time

However, I don't think the golang code will cache those internally, so maybe a small improvement if we cached those for the http -> https redirect we are expecting when we start at http, avoiding a roundtrip to the system resolver even if the value is cached there.

Interesting idea. Could be a quick win

Yeah. I am not sure how to solve that one. Iterative process of doing one level at a time is a lot of wrapping/stitching together. I guess you could try to embed something like ZDNS inside (and maybe this is why there has been work to make it more library work) to do faster lookups than what we are doing today

In this case, are you describing making the same number of DNS requests, but with more optimized code? If so, I would be skeptical of that, mainly because the delay is probably i/o bound. For my dataset, I’m aware of a large number of domains that have very high latency DNS servers. I may be misunderstanding what you’re describing, though

I wonder if part of it is also redirects like domain.com -> www.domain.com, where that doesn't take advantage of multiplexing. I wonder if you could be a little 'clever' and try to get the roundtripper to maintain a connection per IP address, rather than per a hostname.

That could work, as long as it’s HTTP and not HTTPS. It will break on some (possibly many) of the TLS >= 1.2 cases, because of SNI, which will require a new TLS negotiation with the appropriate/new name

Have you considered globally caching DNS across all of the senders? Maybe that’s what you were describing with your idea about zdns “embedding”?

A global cache was one if my first thoughts, but I immediately started arguing points against myself

First, it may help me, but it wouldn’t help all target sets. Only those that have a lot of repeated names over the duration of the session, obviously

In practice, my data tends to have a lot of those, because all of my targets are ultimately assets under the same organization/entity. So I see a lot of cases like:

1. “Parked” names / defensive registrations. Like “thecompanyname.com”, “mycompanyname.com”, “thecompanynameisbad.com”,“companyname2024.com”, … which all may redirect to “www.companyname.com”. In some cases there are literally hundreds of names that redirect to one site. And many such sites.
2. Sites with redirect-based SSO. Like “app.companyname.com”, “app2.companyname.com”, “app3.companyname.com”, … all redirecting to sso.companyname.com

These “feel” very common for my target list but I have to look at the data to gauge the magnitude, maybe by adding simple logging of all DNS requests and doing a uniq -c to see how many repeated. I could also do a wire capture, I guess

Caching globally would also have sharply diminishing returns for those whose target sets don’t consist primarily of targets with the same organizational “owner”, though. Which may be most users

Caching globally might also have potential locking issues impacting performance, if the requests were performed by each sender, but there must be some way to avoid a lock for every new lookup, or reduce the contention. It wouldn’t be an issue if the DNS was a dedicated thread, though

Just thinking out loud here, I haven’t really given any of this proper thought. And it’s early, I probably misunderstood a few of your thoughts- hopefully not all of them 😊

Very happy to have you thinking about this, even though our use patterns may differ quite a bit. I’ve contributed a lot of bug fixes and minor features, but my golang experience and knowledge is sorely lacking. I have to defer to others on any non-trivial implementations, or anything too deep into the core, which is frustrating. I’m used to being more “hands-on” helpful

Anyway, thanks for the discussion, let me know if there’s anything I can do to help in the way of profiling or more primitive data collection