From: Brian Warner Date: Thu, 6 Sep 2007 01:16:21 +0000 (-0700) Subject: provisioning.py: add file/server availability numbers X-Git-Url: https://git.rkrishnan.org/architecture.txt?a=commitdiff_plain;h=849dbb4950e2d0c09a07ab67d69aaa9659501282;p=tahoe-lafs%2Ftahoe-lafs.git provisioning.py: add file/server availability numbers --- diff --git a/src/allmydata/provisioning.py b/src/allmydata/provisioning.py index fdfb582b..a666dbcf 100644 --- a/src/allmydata/provisioning.py +++ b/src/allmydata/provisioning.py @@ -1,10 +1,43 @@ from nevow import inevow, loaders, rend, tags as T from twisted.python import util +import math def getxmlfile(name): return loaders.xmlfile(util.sibpath(__file__, "web/%s" % name)) +# factorial and binomial copied from +# http://mail.python.org/pipermail/python-list/2007-April/435718.html + +def factorial(n): + """factorial(n): return the factorial of the integer n. + factorial(0) = 1 + factorial(n) with n<0 is -factorial(abs(n)) + """ + result = 1 + for i in xrange(1, abs(n)+1): + result *= i + if n >= 0: + return result + else: + return -result + +def binomial(n, k): + if not 0 <= k <= n: + return 0 + if k == 0 or k == n: + return 1 + # calculate n!/k! as one product, avoiding factors that + # just get canceled + P = k+1 + for i in xrange(k+2, n+1): + P *= i + # if you are paranoid: + # C, rem = divmod(P, factorial(n-k)) + # assert rem == 0 + # return C + return P//factorial(n-k) + class ProvisioningTool(rend.Page): addSlash = True docFactory = getxmlfile("provisioning.xhtml") @@ -132,8 +165,10 @@ class ProvisioningTool(rend.Page): " no convergence)", i_sharing_ratio) # Encoding parameters - encoding_choices = [("3-of-10", "3-of-10"), - ("25-of-100", "25-of-100"), + encoding_choices = [("3-of-10", "3.3x (3-of-10)"), + ("5-of-10", "2x (5-of-10)"), + ("8-of-10", "1.25x (8-of-10)"), + ("25-of-100", "4x (25-of-100)"), ] encoding_parameters, i_encoding_parameters = \ get_and_set("encoding_parameters", @@ -161,6 +196,20 @@ class ProvisioningTool(rend.Page): add_input("Servers", "How many servers are there?", i_num_servers) + # availability is measured in dBA = -dBF, where 0dBF is 100% failure, + # 10dBF is 10% failure, 20dBF is 1% failure, etc + server_dBA_choices = [ (20, "99% [20dBA] (14min/day or 3.5days/year)"), + (30, "99.9% [30dBA] (87sec/day or 9hours/year)"), + (40, "99.99% [40dBA] (60sec/week or 53min/year)"), + (50, "99.999% [50dBA] (5min per year)"), + ] + server_dBA, i_server_availability = \ + get_and_set("server_availability", + server_dBA_choices, + 20, int) + add_input("Servers", + "What is the server availability?", i_server_availability) + # deletion/gc/ownership mode ownership_choices = [ ("A", "no deletion, no gc, no owners"), ("B", "deletion, no gc, no owners"), @@ -429,6 +478,19 @@ class ProvisioningTool(rend.Page): (100.0 * share_data_per_server / share_space_per_server)]) + # availability + file_dBA = self.file_availability(k, n, server_dBA) + user_files_dBA = self.many_files_availability(file_dBA, + files_per_user) + all_files_dBA = self.many_files_availability(file_dBA, total_files) + add_output("Users", + T.div["availability of: ", + "arbitrary file = %d dBA, " % file_dBA, + "all files of user1 = %d dBA, " % user_files_dBA, + "all files in grid = %d dBA" % all_files_dBA, + ], + ) + all_sections = [] all_sections.append(build_section("Users")) @@ -449,3 +511,53 @@ class ProvisioningTool(rend.Page): ] return f + + def file_availability(self, k, n, server_dBA): + """ + The full formula for the availability of a specific file is:: + + 1 - sum([choose(N,i) * p**i * (1-p)**(N-i)] for i in range(k)]) + + Where choose(N,i) = N! / ( i! * (N-i)! ) . Note that each term of + this summation is the probability that there are exactly 'i' servers + available, and what we're doing is adding up the cases where i is too + low. + + This is a nuisance to calculate at all accurately, especially once N + gets large, and when p is close to unity. So we make an engineering + approximation: if (1-p) is very small, then each [i] term is much + larger than the [i-1] term, and the sum is dominated by the i=k-1 + term. This only works for (1-p) < 10%, and when the choose() function + doesn't rise fast enough to compensate. For high-expansion encodings + (3-of-10, 25-of-100), the choose() function is rising at the same + time as the (1-p)**(N-i) term, so that's not an issue. For + low-expansion encodings (7-of-10, 75-of-100) the two values are + moving in opposite directions, so more care must be taken. + + Note that the p**i term has only a minor effect as long as (1-p)*N is + small, and even then the effect is attenuated by the 1-p term. + """ + + assert server_dBA > 9 # >=90% availability to use the approximation + factor = binomial(n, k-1) + factor_dBA = 10 * math.log10(factor) + exponent = n - k + 1 + file_dBA = server_dBA * exponent - factor_dBA + return file_dBA + + def many_files_availability(self, file_dBA, num_files): + """The probability that 'num_files' independent bernoulli trials will + succeed (i.e. we can recover all files in the grid at any given + moment) is p**num_files . Since p is close to unity, we express in p + in dBA instead, so we can get useful precision on q (=1-p), and then + the formula becomes:: + + P_some_files_unavailable = 1 - (1 - q)**num_files + + That (1-q)**n expands with the usual binomial sequence, 1 - nq + + Xq**2 ... + Xq**n . We use the same approximation as before, since we + know q is close to zero, and we get to ignore all the terms past -nq. + """ + + many_files_dBA = file_dBA - 10 * math.log10(num_files) + return many_files_dBA