virtualize all the things... ?
-
@bj said in virtualize all the things... ?:
Specifically, do you virtualize your sip servers and heavy disk IO servers?
Definitely. If these are having issues, something else is wrong. At @NTG we've been 100% virtual for VoIP / SIP since 2005 at least.
-
@bj said in virtualize all the things... ?:
So, my questions: is there ever a time when you do not want to virtualize a server?
Essentially, no. There are times, but they are so uncommon and special case that it is better, IMHO, to assume that they don't exist. Even at the world's biggest banks where they do this, it represents less than .001% of their workloads.
-
The world's biggest disk IO and database systems are all virtual. It's only in the SMB space that having physical systems of this nature are even an inkling of an idea. The giant "billions and billions of daily transaction" systems are all virtual.
-
@scottalanmiller
Since we always should be using virtualization, is there anything wrong with having only one VM that requires a lot of resources from the host? -
Facebook runs all physical, but that's an edge case.
-
@black3dynamite said in virtualize all the things... ?:
@scottalanmiller
Since we always should be using virtualization, is there anything wrong with having only one VM that requires a lot of resources from the host?One to one is just fine. Nothing in the rule to "always virtualize" implies that we always consolidate. Totally different considerations.
-
@marcinozga said in virtualize all the things... ?:
Facebook runs all physical, but that's an edge case.
Is that still true? That was news based on them using crappy Atom servers six or seven years ago. FB is arguably virtualizing in a macro way, though, they do cluster level virtualization. Same as we did on Wall St. for massive decision clusters.
-
@scottalanmiller said in virtualize all the things... ?:
@marcinozga said in virtualize all the things... ?:
Facebook runs all physical, but that's an edge case.
Is that still true? That was news based on them using crappy Atom servers six or seven years ago. FB is arguably virtualizing in a macro way, though, they do cluster level virtualization. Same as we did on Wall St. for massive decision clusters.
I have no idea, but their reasoning was they didn't have underutilized servers so there was no point virtualizing.
-
@marcinozga said in virtualize all the things... ?:
@scottalanmiller said in virtualize all the things... ?:
@marcinozga said in virtualize all the things... ?:
Facebook runs all physical, but that's an edge case.
Is that still true? That was news based on them using crappy Atom servers six or seven years ago. FB is arguably virtualizing in a macro way, though, they do cluster level virtualization. Same as we did on Wall St. for massive decision clusters.
I have no idea, but their reasoning was they didn't have underutilized servers so there was no point virtualizing.
Which is no logic at all as that's the reason not to consolidate and has nothing to do with virtualization. This suggests that a confused software guy was quoted and not someone from even the IT department. All it tells us is that the guy being quoted doesn't know what virtualization is and made up something to be a sound bite.
-
What's your favorite open source virtualization platform for SMB?
-
@bj said in virtualize all the things... ?:
What's your favorite open source virtualization platform for SMB?
Don't have one. Okay, okay, I do. I like Xen. Mostly because I've been on it since like 2003. I like PV conceptually. Xen has always treated me well.
KVM was a silly project and was designed to splinter the market, so I dislike it because of that. Today, Xen and KVM are sibling products of the Linux Foundation. In reality, KVM is the better choice most of the time.
-
@marcinozga said in virtualize all the things... ?:
they didn't have underutilized servers
I seriously doubt that
-
@wirestyle22 At what point do you consider a server not "under utilized"?
-
@marcinozga said in virtualize all the things... ?:
I have no idea, but their reasoning was they didn't have underutilized servers so there was no point virtualizing.
As @wirestyle22 points out... this is almost certainly just lying. Getting good utilization from a single server without virtualization is hard enough. Doing it with all servers is absurdly impossible. That they had more than one server at all suggests that their entire purchasing strategy is likely based around not consolidating (e.g. they bought servers sized to make them appear utilized, rather than fewer larger servers that could consolidate.)
If your goal is to make virtualization seem unnecessary by first conflating the value with that of consolidation and then sizing servers to be heavily used and you have people who don't catch on to either part, it can seem plausible. But it requires a lot of work and wasted money and in the end, if we looked at what they have, I bet the proof that they needed to virtualize is pretty evident.
-
@bj said in virtualize all the things... ?:
@wirestyle22 At what point do you consider a server not "under utilized"?
That is actually my point
-
@bj said in virtualize all the things... ?:
@wirestyle22 At what point do you consider a server not "under utilized"?
When the CPU is 100%
Sounds silly, but that's really kind of the answer.
-
@scottalanmiller said in virtualize all the things... ?:
@bj said in virtualize all the things... ?:
@wirestyle22 At what point do you consider a server not "under utilized"?
When the CPU is 100%
Sounds silly, but that's really kind of the answer.
Right
-
If you are hitting 100%, I presume you then back off a little? That seems... unhealthy in the long term.
-
@bj said in virtualize all the things... ?:
If you are hitting 100%, I presume you then back off a little? That seems... unhealthy in the long term.
Any unused resource is not being utilized, making a server underutilized. That's the point we are making.
-
I understand the point, and on that point you are absolutely correct. My question is in practice, are you actually trying to hit that 100% mark? It seems like the services would do well to have a little buffer room in there, considering spikes in load and all. But maybe we're talking apples and oranges here. I'm thing production web / database servers, where speed is a priority. Maybe you are considering a less sensitive work load?
